Hermosa Research·Descriptive
AI Coding-Agent Primitives
A 1,708-cell comparison of 61 coding-agent capabilities across 28 harnesses. See how tools implement plan mode, subagents, hooks, memory, MCP, and checkpoints as commands, modes, prompts, or extensions.
- Published
- 1 June 2026
- Data as of
- 7 September 2026
- Research type
- [descriptive]Maps or measures a landscape, no hypothesis under test. The rigor lives in scope, selection criteria, and provenance.[tentative]Reasonably supported, but parts are still unverified or liable to drift.
- Topics
- [ai-agents][developer-tools][comparison]
Method: We coded 1,708 cells from first-party documentation and product interfaces for 28 selected harnesses. Each cell preserves the documented mechanism; 244 remain unconfirmed as of 2026-08-22.
Background
Most coding agents can plan, edit files, run commands, call tools, and retain context. The difficult comparison is not whether a capability exists, but how a developer invokes it. Plan mode may be a command in one tool, a keyboard toggle in another, and an extension elsewhere.
This matrix records those mechanisms directly. It covers 61 capabilities across 28 selected harnesses and retains the documented label for every cell. It is a reference, not a ranking. The wider catalogue contains 106 coding agents, including 78 that have not yet been coded into the matrix.
Key findings
- 1Interfaces differ more than capabilities. Planning, memory, delegation, and tool access recur across tools, but they appear as commands, keyboard modes, prompts, or extensions.
- 2Plan mode is usually a mode, not a /plan command. Codex, Qwen Code, and Copilot document a command. Claude Code, Cursor, opencode, Cline, Roo, and Kilo expose planning through a mode or interface control.
- 3Persistent goals remain uncommon. The matrix confirms a native persistent objective in Codex and Qwen Code. Third-party references to Claude Code /goal are not supported by Anthropic’s official documentation.
- 4Extensions standardise workflows across tools. Cross-tool packs such as GitHub Spec Kit carry the same specification, planning, task, and implementation sequence into otherwise different harnesses.
- 5Unknown is a data state, not a negative result. 228 of 1,708 cells remain unconfirmed. They are shown as “?” rather than being counted as unsupported.
The support matrix
The matrix covers 28 harnesses and 61 primitives. Each cell preserves the documented mechanism and maps it to a support level. Filter the rows, hide harnesses, collapse categories, or download the source data.
Harnesses (28/28)
| Primitive |  Claude Code |  Codex |  opencode |  Aider |  Pi |  Amp |  Cline |  Kilo |  Cursor |  Devin Desktop |  Zed |  Copilot |  Devin |  Jules |  Kiro |  Junie |  Factory Droid |  Augment Code |  OpenHands |  Trae |  Warp |  CodeBuddy |  Mistral Vibe |  Roo | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Plan modeorigin: Cline | Shift+Tab | /plan | Tab | /architect | extension (Shift+Tab pkg) | /plan | Deep mode | prompt | prompt | Plan/Act | Architect mode | Shift+Tab | plan preview | prompt | /plan, plan agent | plan-first | /plan slash command | plan-first | Shift+Tab (CLI); Quick Plan (specs) | /plan, --plan flag, Shift+Tab | Spec mode (Shift+Tab; --use-spec) | ? | prompt | SOLO (plan before execute) | /plan | ? | plan agent (Shift+Tab) | Architect mode |
| Spec / specifyorigin: GitHub Spec Kit | – | – | – | – | SPEC mode (pkg) | – | – | – | – | – | – | – | – | – | – | – | ? | ? | Specs (requirements.md + design.md + tasks.md) | ? | Spec Mode (--use-spec; --spec-model) | ? | ? | Spec workflow (/spec command; generates spec.md + tasks.md + checklist.md in .trae/specs/, version-controlled) | ? | ? | ? | – |
| Constitution / principlesorigin: GitHub Spec Kit | – | – | – | – | – | – | – | – | – | – | – | – | – | – | – | – | ? | ? | Steering files (.kiro/steering/*.md) | ? | ? | ? | ? | ? | AGENTS.md / WARP.md | ? | ? | – |
| Brainstorm / discoveryorigin: (generic) | conversational | conversational | conversational | /ask | conversational | conversational | conversational | conversational | conversational | Plan | Ask mode | ask | conversational | conversational | explore agent | conversational | Ask Devin (codebase Q&A before session) | conversational | conversational (vibe sessions) | Ask mode (read-only exploration) | conversational (Missions planning conversation) | Quick Ask (read-only) | conversational | conversational (Chat mode Q&A) | conversational | Ask mode | ? | Ask mode |
| Deep research modeorigin: Amp | – | – | – | – | – | research | Deep mode | – | – | – | – | – | – | – | /research agent | – | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | – |
| Clarifying questionsorigin: Cline | yes (plan) | yes | yes | limited | extension | yes | yes | yes | limited | Plan | Architect | yes (plan) | yes | limited | yes | yes | yes (Ask Devin; !ask in Slack) | yes (plan) | yes (spec phases + Quick Plan) | yes (plan mode; approves plan before implementation) | yes (Spec Mode: read-only exploration + clarifying questions) | ? | yes | ? | Ask questions permission (Always/Conditional/Never) | yes (Plan; AskUserQuestion) | ask_user_question | Architect |
| Estimate / complexityorigin: SuperClaude | – | – | – | – | – | – | – | – | – | – | – | – | – | – | – | – | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | – |
| Act / implementorigin: Cline (named split) | native | native | Build mode | /code | native | native | native | native | native | Act | Code mode | Agent mode | Cascade | native | agent | agent | native (autonomous cloud session) | native | Autopilot (default autonomous) + Supervised mode | Code mode (autonomous multi-step plan + execution) | native (Auto mode) | native (Agent) | native | Builder mode | Agent Mode (local + cloud) | Craft mode | native (default agent) | Code mode |
| Edit / apply difforigin: Aider (diff edits) | native | native | native | diffs | Edit tool | native | native | native | native | native | native | native | diff-staging | native | native | native | native | native | native (per-hunk in Supervised; full diff view in Autopilot) | native (diff viewer before accept) | native (diff shown; wait for approval) | native (diff review per action) | native | native (Accept/Reject All; ^Y/^N) | Interactive Code Review panel | native (Edit/MultiEdit; diff review) | write_file / search_replace | native |
| Run terminal commandsorigin: (generic) | yes | yes | yes | yes | Bash tool | yes | yes | yes | yes | yes | yes | yes | Turbo mode | yes | yes | yes | yes (shell terminal in VM) | yes | yes (approve/reject/run/trust per command) | yes (! prefix; brave mode controls approval) | yes (risk-gated by autonomy) | yes (terminal tool) | yes | yes (Builder runs Shell; Chat one-click) | yes (configurable approval per Agent Profile) | yes (Bash; permission-gated) | bash | yes |
| Debugorigin: Roo Code (Debug mode) | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | Debug mode | /debug | prompt | prompt | prompt | prompt | prompt | prompt | Bugfix Specs (bugfix.md + design.md + tasks.md) | /debug mode (live IDE debugger, breakpoints, step) | prompt (droid exec natural-language debugging) | ? | prompt | TRAE-debugger built-in skill (starts a debugging server, collects logs over HTTP; hypothesis -> instrumentation -> reproduction -> analysis -> fix -> verification) | agent debug (autonomous bug investigation; reproduces issue, inspects likely failure paths) | ? | prompt | Debug mode |
| Test generationorigin: (generic) | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | task agent | Workflow | prompt | task agent | prompt | /test slash command; video test recordings | prompt | Autopilot test generation (multi-file); property-based tests | ? | prompt (droid exec natural-language test fixing/generation) | ? | ? | ? | ? | ? | prompt | prompt |
| Code revieworigin: (parallel) | /review | /review | /review | manual | extension | AI review | Oracle subagent | prompt | prompt | – | – | BugBot | review Workflow | prompt | rubber-duck agent | via skills | Devin Review (PR bot + /review slash command) | Critic agent (auto peer-review) | Supervised mode per-hunk review; BugBot via hooks | /review (local diff); GitHub Action (PR inline comments) | /review | native GitHub PR review ("auggie review") | GitHub Action / prebuilt automation | TRAE-code-review built-in skill (reviews merge requests/diffs; structured feedback on quality, correctness, best practices) | Code Review panel + Interactive Code Review | /security-review | prompt | – |
| Refactororigin: (generic) | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | prompt | via modes | prompt | refactor Workflow | prompt | prompt | prompt | prompt | prompt | prompt (chat or spec task) | ? | prompt (droid exec natural-language refactor; --use-spec plans first) | ? | ? | ? | agent refactor (write/refactor features via autonomous agents) | ? | prompt | via modes |
| Goal (persistent objective)origin: Codex | unconfirmed* | /goal (flag) | – | – | – | /goal | – | – | – | – | – | non-persistent | – | – | non-persistent | non-persistent | ? | – | /goal (iterative loop until verified) | ? | Missions (describe goal; orchestrator manages work to completion) | ? | convo-goal (SDK) | ? | cloud agent task (runs to completion) | /goal | ? | – |
| Orchestrator / boomerangorigin: Roo Code (Boomerang) | subagents | – | – | – | TEAM/CHAIN/PIPELINE (pkg) | subagents | subagents | recipes | – | – | Orchestrator mode | subagents | – | Parallel Agents | Fleet mode | Agent Manager | Managed Devins (coordinator delegates to child sessions) | ? | subagents (main agent delegates to built-in + custom sub-agents) | subagents (auto-delegated subtasks) | Mission Mode (multi-agent orchestration) | automatic subagent delegation (Agent detects and offers to delegate to a subagent) | agent delegation (SDK) | main "Agent" auto-delegates a task to a subagent when the subagent's description is highly relevant to that task | Warp Agent (cross-harness subagent orchestration via Oz) | subagents (auto-delegated) | subagents via task tool | Orchestrator mode |
| Subagents (own context)origin: Claude Code | /agents | agents | agents | – | extension (Fleets/Crews) | fork subagent | Librarian/Oracle | yes | – | Subagents | (modes) | subagents (10/50) | – | yes | subagents | subagents | Managed Devins + CLI subagents (foreground/background) | – | yes (custom .kiro/agents/*.md with isolated context) | .junie/agents/ Markdown+YAML files, isolated context | Custom Droids (.factory/droids/*.md; fresh context) | yes (.augment/agents/*.md; isolated context) | yes | Subagents (.trae/agents/ project scope; Markdown+YAML frontmatter with name/description/model/tools/mcpServers; independent context windows whose intermediate reasoning does not pollute the main Agent's chat history; main Agent auto-delegates by matching description to intent) | cross-harness subagents (Oz) | subagents (.codebuddy/agents/) | subagents (.vibe/agents/*.toml; task tool) | (modes) |
| Parallel agents / swarmorigin: Roo/Kilo + Copilot Fleet | via plugins | – | – | – | extension | parallelism | yes | – | – | – | up to 10/50 | Fleet mode | cloud (Devin) | Parallel Agents | Fleet mode | yes (default) | yes, parallel cloud sessions via Managed Devins | up to 60 concurrent tasks | yes (subagents run in parallel) | GitLab CI/CD multi-task; worktree isolation | Mission Mode workers + worktree parallelism | yes (subagents run in parallel; Remote Agents run concurrently in cloud) | yes | ? | parallel cloud subagents (Oz); local parallel tabs + worktrees | background agents | ? | up to 10/50 |
| Task breakdown / to-dosorigin: (generic) | TODO list | skills | agents | – | extension | task lists | yes | yes | yes | task mgmt | mode split | to-dos | yes | threads | task lists (synced) | task lists (synced) | Playbooks + /plan | plan steps | tasks.md (spec phase 3, wave-based parallel execution) | ? | Missions milestones (structured plan; validation per milestone) | Task Manager (auto-created tasks; persists under ~/.augment) | TaskTrackerTool (available built-in tool; added explicitly in agent config) | SOLO auto task breakdown | Agent Task Lists (auto-generated) | /todos (Plan checklist) | todo list (Ctrl+T to toggle view) | mode split |
| Dynamic workflow (scripted)origin: Cline (Workflows) | no native | – | – | – | CHAIN/PIPELINE (YAML) | custom workflows | reusable workflows | recipes | – | Workflows | custom modes | plugins | Workflows | – | orchestration | orchestration | Playbooks (reusable multi-step task templates) | ? | Specs (3-phase pipeline: requirements -> design -> tasks) | ? | ? | ? | ? | ? | YAML Workflows + Warp Drive Prompts | ? | ? | custom modes |
| Autonomous looporigin: (community) | – | – | – | – | extension | /goal autonomous | Deep mode | yes | – | Focus Chain | – | Background Agent | Cascade (Turbo) | – | coding agent | Managed Agents | yes (async cloud sessions run until done; Automations) | yes | Autopilot mode + /goal iterative loops | Brave Mode (runs all sensitive actions without per-step approval) | Autonomy Level (High) + Mission Mode (work without per-step approval) | Agent Auto (edits files, runs terminal commands, uses MCP tools automatically) | yes | SOLO mode (autonomous after plan) | YOLO mode (all permissions set to Always allow) | /goal (until condition met) | auto-approve agent | – |
| Background / async (cloud)origin: (Devin-style) | no (local) | Codex cloud | – | – | – | background tasks | – | – | – | Cline cloud | Kilo cloud | Background Agent | handoff to Devin | – | /delegate agent | Managed Agents API | yes, native; all sessions are async cloud | cloud VM async | Kiro Web autonomous mode (cloud sandbox + PR) | GitHub Actions / GitLab CI runners (on-infra cloud exec) | Droid Computers (persistent cloud dev environment) | Remote Agents (cloud) | automations (Agent Canvas cloud) | ? | cloud agents on Oz (background tasks) | ? | ? | Kilo cloud |
| Scheduled / cron tasksorigin: Qwen Code | – | – | – | – | extension | cron scheduled | – | – | – | – | – | – | – | – | – | – | Scheduled sessions (cron + one-time); Automations | Scheduled Tasks (daily/weekly/monthly) | Automations (hourly/daily/CRON schedules, up to 5 per automation) | ? | Automations (Scheduled trigger; natural-language cadence or cron) | Cosmos scheduled automations (5-field cron expression, timezone-aware) | automations (cron schedule) | ? | Scheduled Agents (cron syntax via Ambient Agents) | ? | /loop (recurring prompt) | – |
| Worktree isolationorigin: Cline | via plugins | – | – | – | extension | worktree | – | – | – | Worktrees | Kilo cloud | Worktrees (3.2) | – | – | Worktree isolation | ? | ? | cloud VM per task | cloud sandboxes (Web; each run is isolated) | /worktree (git worktrees, parallel sessions) | -w, --worktree [name] | isolated cloud env per Remote Agent | ? | ? | isolated worktrees (parallel subagents each run in their own worktree) | ? | --worktree NAME (creates/reuses a git worktree, runs session inside it) | Kilo cloud |
| Project init (memory gen)origin: Claude Code (/init) | /init | /init | /init, /init-deep | – | extension | init | yes | – | – | init memory bank | init | rules gen | yes | – | yes | GEMINI.md | ? | – | Generate Steering Docs (.kiro/steering/) | auto-imports from other agents' memory files | /init | ? | yes | ? | /init (generates AGENTS.md) | /init | ? | init |
| Persistent memory fileorigin: CLAUDE.md / AGENTS.md / Memory Bank | CLAUDE.md | AGENTS.md | AGENTS.md | conf file | AGENTS.md | cross-session memory | persistent threads | yes | yes | Memory Bank | Memory Bank | rules + AGENT.md | workspace memory | threads | AGENTS.md | GEMINI.md / .agents | AGENTS.md + Knowledge (org-scoped knowledge base) | AGENTS.md | .kiro/steering/*.md (workspace) + ~/.kiro/steering/ (global) | AGENTS.md (.junie/AGENTS.md or project root) | AGENTS.md + .factory/memories.md | Memories (auto-generated) + AGENTS.md/.augment/rules | AGENTS.md / Skills | AGENTS.md + .trae/project_rules.md + .trae/user_rules.md | AGENTS.md + Agent Memory (Oz cloud, cross-session) | CODEBUDDY.md | AGENTS.md (~/.vibe/ + project) | Memory Bank |
| Memory edit / updateorigin: Claude Code (/memory) | /memory | edit AGENTS.md | edit file | – | edit AGENTS.md | update | yes | yes | ? | update memory bank | update | edit rules | edit memory | Rules Library (edit .rules) | edit instructions | edit GEMINI.md | Knowledge CRUD (UI + API); Devin auto-suggests Knowledge | Memory System (repo-level) | edit steering files; Generate Steering Docs command | edit AGENTS.md (project or global ~/.junie/AGENTS.md) | ? | Memory Review (approve/edit/remove in chat panel) | edit AGENTS.md / Skills | edit user_profile.md / project_memory.md directly, or via chat (AI auto-creates/updates/deletes memories on request) | auto-extracted from transcripts by Oz (no manual edit command) | /memory; Auto Memory | edit AGENTS.md | update |
| Codebase indexing / RAGorigin: Aider (repo map) / Cursor | grep-based | grep-based | yes | repo map | grep/bash | yes | Sourcegraph search | yes | LSP-enhanced | codebase index | codebase indexing | index | repo-scale | LSP | code intelligence | yes | DeepWiki (auto-indexed repo wikis + semantic search) | yes | codebase indexing (#codebase context provider) | semantic indexing (embeddings, checkbox in IDE plugin) | grep/glob-based (Read, LS, Grep, Glob tools) | Context Engine (semantic indexing) | grep/glob-based | codebase indexing (auto <=5k files) | Codebase Context (local embeddings-based indexing) | grep/glob-based | grep (ripgrep) | codebase indexing |
| Context compactionorigin: Claude Code (/compact) | /compact | auto | /compact | /tokens | token-efficient core | ToolSearch on demand | yes | yes | ? | context condensing | context condensing | auto | yes | auto + /compact | token mgmt | auto (~135k tokens) | /compact (CLI) | ? | /compact (auto on overflow) | ? | /context (window usage breakdown) | ? | LLMSummarizingCondenser | auto-triggered once context exceeds the allowed window during task processing, or manual Compact button (built-in "Agent" + other agents with Max mode enabled) | /compact; /fork-and-compact | /compact | /compact | context condensing |
| Clear / reset contextorigin: (generic) | /clear | /new | /new | /clear, /reset | /new | new | new thread | /clear | ? | new task | new task | new chat | ? | new thread | /clear | ? | /clear or /new (CLI); sleep/archive (Slack) | new task | /clear; /context clear; /chat new | /new (clear context + optional prompt) | /clear | New chat icon (Chat panel) | /new | new chat (clears the previous context) | /new | /clear | /clear | new task |
| Checkpoints / rollbackorigin: Cline (per-tool-call) | git-based | git-based | /undo, /redo | /undo | git checkpointing (pkg) | rewind | ? | ? | ? | per-tool-call | per-call | checkpoints / restore | revert to step (hover prompt + revert arrow) + named snapshots | Restore Checkpoint | git-based rewind snapshot per prompt (Esc Esc / /undo / /rewind) | sandbox | snapshot-per-session (each session boots clean snapshot) | ? | /rewind (CLI) + Revert All Changes (Autopilot) | roll back all changes (decline task changes) | /rewind-conversation | checkpoints (auto snapshots; revert) | conversation fork (SDK) / git-based | ? | ? | /rewind (per-prompt checkpoints) | /rewind (Alt+Up / Ctrl+P) | per-call |
| Intent / activity trackingorigin: Windsurf | – | – | – | – | – | – | – | – | – | – | – | – | Cascade tracking | – | – | – | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | – |
| Resume / continueorigin: (generic) | /resume | /resume | /sessions | --restore | session rewind | resume | threads | yes | yes | task history | sessions | resume | yes | threads | session history | synced sessions | yes, /resume <id> (CLI); cloud session resume; CLI /handoff | yes | sessions (Web + iOS sync across platforms) | /history (last 10 sessions); --session-id flag | droid --resume [sessionId]; droid exec -s <id> | --continue / --resume (persistent session state) | --resume / --last | ? | cloud-synced conversation history (CMD+Y); cross-device | /resume; --continue / --resume | /resume; --continue/-c; --resume SESSION_ID | sessions |
| Rewind / branch from messageorigin: Pi / Qwen Code | – | /fork | – | – | rewind + branch | message rewind | – | – | – | – | – | – | – | – | – | – | ? | ? | /rewind (fork at earlier turn) | ? | Fork a session | ? | conversation fork (SDK) | ? | Conversation Forking (/fork-from, right-click block) | /rewind; /branch (/fork) | /rewind | – |
| Steering / interrupt mid-runorigin: Pi | esc | ? | ? | ? | Enter=steer, Alt+Enter=queue | Ctrl+C (cancel) | ? | Ctrl+C (interrupt) | esc | ? | ? | ? | ? | stop button / Send Now | steer (enter new prompt while working) | comment on Artifact (live) | yes, message session mid-run; mute/unmute in Slack | pause button + chat feedback | Queue Steering (CLI v2.7+); Interrupt Execution button | follow-up prompts while agent is working | Ctrl+C (interrupt; twice to exit) | queued follow-up messages while agent is working (Cosmos sessions) | Esc (CLI) / send_message while running (SDK) | ? | mid-run steer input; Remote Control | ? | Escape | ? |
| Session share / exportorigin: opencode (/share) / Amp | /copy, export | /copy | /share, export | – | share session | /export | shared threads | session export (md/json/yaml) | ? | – | – | – | – | channels | /share file (markdown); /share gist | Agent Manager | session link sharing; #devin-runs Slack channel | ? | export (CLI v2.6+: markdown/JSON transcript export) | ? | /share (with org); /copy (session) | Share dialog (grant a teammate viewer access, copy link, revoke access) | trajectory export (JSON) | ? | Session Sharing (link, real-time sync, web/mobile view) | /export | ? | – |
| Model switchingorigin: (generic) | /model | /model | /models | /model | /login, config (20+) | config | auto-select | multi-model | yes | profiles | profiles | model picker | model routing | model routing | /model (multi) | model picker | /model (CLI); !fast / !ultra / !lite (Slack); adaptive routing | tier-based (Gemini 3 Flash / Pro) | /model (CLI); model picker (IDE/Web); persistent preferences | /model (10+ models, effort levels, BYOK) | /model; -m, --model <id>; Ctrl+N; BYOK | /model + --model (Claude/GPT/Gemini/Kimi) | /model or profile selector | model picker; /switch model:<name>; BYOK | model picker dropdown; /model; BYOK | /model; --model | /model | profiles |
| Cost / token trackingorigin: Aider / Claude Code | /cost | usage | yes | /tokens | token display | real-time usage | leaderboards | ? | ? | cost display | cost | usage | credits | token usage display | usage | Quota screen (per-model bars) | ACU (Agent Compute Unit) usage per session; Session Insights | ? | /usage (CLI); subscription dashboard (IDE: credits used/remaining) | /usage (token breakdown + balance per session) | /cost; /stats [period] | Usage Dashboard (My Usage / Organization Usage; shown in $ or credits) | yes (metrics API) | Dollar Usage display (used + remaining) for the current subscription period, in the billing/subscription panel | per-turn credit chip + Settings > Billing; Enterprise Analytics API | /cost; /stats | --max-price (indicative only) | cost |
| Print / headless modeorigin: (generic) | -p / --print | exec | run | --message | print mode | --prompt / -p (--output-format text|json|stream-json) | CLI pipe | headless | crush run (--quiet) | – | – | CLI | – | – | --prompt | agy CLI | devin -p "prompt" (CLI single-turn stdout) | jules remote new (CLI) | --no-interactive (CLI headless via KIRO_API_KEY) | headless (junie --auth=... "prompt") | droid exec (-o text/json/stream-jsonrpc) | --print (single-shot, non-interactive) | --headless -t / -f | ? | Oz CLI (non-interactive; --output-format json/text) | -p / --print (--output-format json) | --prompt (--output text|json|streaming) | – |
| Custom slash commandsorigin: (generic) | yes (.claude/commands) | yes (prompts) | yes (.opencode/command) | – | prompt templates | custom workflows | yes | yes | yes (custom/saved prompts as Markdown command files) | workflows | yes | yes (prompts) | yes | skills via / (SKILL.md) | yes | inherited | yes, org-level custom slash commands via Settings > Customization | ? | yes (custom agents invokable as slash commands) | .junie/commands/ Markdown files with YAML frontmatter | yes (.factory/commands/*.md; /commands) | yes (.augment/commands/<name>.md) | yes (commands via plugins/skills) | Commands (.trae/commands/ project scope, ~/.trae/commands global scope; invoked via / in chat) | Warp Drive Prompts (appear in / menu) | yes (.codebuddy/commands/*.md) | skills (user-invocable: /skill-name) | yes |
| Custom modesorigin: Roo Code | via agents | – | agents | chat-modes | 6 modes (pkg) | subagents (task-specific prompts + tool scopes; .qwen/fork-profiles) | – | – | – | limited | custom modes | custom agents | – | – | custom agents | custom agents (.agents/agents/ MD + YAML frontmatter: model + tools) | ? | ? | custom agents (config file: tool scope + permissions + context) | subagents (.junie/agents or .agents; tools/disallowedTools, model) | custom droids (.factory/droids/ or ~/.factory/droids/; tool policy + model) | subagents (.augment/agents/ MD; tools allowlist/denylist) | file-based agents (.agents/agents/*.md; tools frontmatter + model) | custom agents (persona prompt; per-agent MCP servers + built-in tools) | Agent Profiles (autonomy + model + tool scope per profile) | ? | agent profiles (.vibe/agents/*.toml) scoped by approval level (auto-approves safe read tools, blocks edits and commands per profile) | custom modes |
| Hooks (event-driven)origin: Claude Code | hooks | limited | limited | – | extension | hooks (PreToolUse, PostToolUse, SessionStart/End, Stop, etc.) | – | – | – | Hooks | – | Pre/PostToolUse / stop | – | – | hooks (8 events) | hooks | Automations (event triggers: Slack, GitHub, Linear, webhooks, cron) | event-driven integrations (webhook triggers) | Agent Hooks (file save/create/delete, prompt submit, PreToolUse, PostToolUse, Stop, spec task) | 7 lifecycle hooks in config.json (SessionStart, PreToolUse, Stop, etc.) | /hooks (lifecycle hooks) | Hooks (PreToolUse, PostToolUse, Stop, SessionStart/End; settings.json) | hooks (6 events) | hooks (lifecycle events; JSON config with global + project scope; multiple project-root hook configs combined at runtime) | ? | hooks (pre/post tool) | hooks (.vibe/hooks.toml: pre_tool, post_tool, post_agent) | – |
| Skills (SKILL.md)origin: Claude Code | Skills | skills | skills | – | Skills (portable) | Skills (SKILL.md in extensions; /skills command) | – | – | – | Skills | – | Agent Skills | – | – | skills | Global / Agent Skills | SKILL.md files in .agents/skills/<name>/SKILL.md | ? | ? | .junie/skills/ SKILL.md + templates (Agent Skills) | Skills (.factory/skills/<name>/SKILL.md; /skills) | Skills (.augment/skills/<name>/SKILL.md; invoked via /skill-name) | Skills (.agents/skills/SKILL.md) | Skills (SKILL.md in .trae/skills/ project or ~/.trae/skills/ global; on-demand loading; also reads .agents/skills/) | Skills (YAML-frontmatter .md files; oz-skills repo) | Skills (.codebuddy/skills/SKILL.md) | Skills (.vibe/skills/<name>/SKILL.md) | – |
| Custom agent profilesorigin: Claude Code | .claude/agents | agents | agents | – | MD + YAML frontmatter | subagent defs | yes | – | – | subagent defs | mode YAML | custom subagents | – | ? | .agent.md | .agents/ | CLI subagent profiles (subagent_explore, subagent_general) | ? | .kiro/agents/*.md (YAML frontmatter: name, description, tools, model) | .junie/agents/ YAML frontmatter (name, tools, model, skills) | .factory/droids/*.md (YAML frontmatter: name, model, tools) | subagent .md files (YAML frontmatter: name, model, tools) | .agents/agents/*.md | custom agents (intelligent gen or manual) | ? | .codebuddy/agents/*.md (YAML frontmatter) | .vibe/agents/*.toml (model, tools, permissions) | mode YAML |
| MCP managementorigin: Anthropic (MCP standard) | /mcp | /mcp | config | limited | extension | MCP mgmt | yes | MCP-first | MCP-extensible | MCP marketplace | yes | .cursor JSON | yes | yes | GitHub MCP | extensions | yes, Devin MCP server + MCP marketplace; CLI MCP config | via MCP | yes (MCP servers tab; /mcp slash command; .kiro/settings/mcp.json) | /mcp + mcp.json (project or user scope); Installation Assistant | /mcp; droid mcp add <name> <url> | Easy MCP (+ OAuth/token); Import from JSON; settings panel | yes (Settings > MCP / mcp.json) | yes (marketplace; .trae/mcp.json; stdio/SSE/HTTP) | MCP servers (per-profile allowlist/denylist; add mid-session) | /mcp; codebuddy mcp add | /mcp (HTTP + stdio) | yes |
| Plugins / marketplaceorigin: Claude Code | /plugin marketplace | skills | plugins | – | npm/git packages | extension mgmt | – | – | – | MCP marketplace | marketplace | Marketplace / /add-plugin | – | – | plugins | plugins | ? | ? | ? | /extensions (git repos, local dirs, HTTP manifests; JetBrains marketplace) | droid plugin install / droid plugin marketplace | Plugins + Marketplaces (auggie plugin marketplace add; /plugins browser) | plugins (GitHub registry) | Extension Store + VS Code Marketplace compatibility + local .vsix install | ? | /plugin (marketplace) | ? | marketplace |
| Rules / instructions filesorigin: Cursor (.cursorrules) | CLAUDE.md | AGENTS.md | AGENTS.md | CONVENTIONS.md | AGENTS.md | QWEN.md | AGENT.md | .goosehints | CRUSH.md / AGENTS.md | .clinerules | .roo/rules | .cursor/rules | rules | .rules | .github instructions | GEMINI.md | AGENTS.md + Knowledge items | AGENTS.md | .kiro/steering/*.md + AGENTS.md (auto-included) | AGENTS.md (.junie/AGENTS.md); .aiignore for file restrictions | AGENTS.md + .factory/rules/ | .augment/rules + AGENTS.md/CLAUDE.md (.augment-guidelines) | AGENTS.md / .openhands/ | AGENTS.md + .trae/rules/ (project & user) | AGENTS.md / WARP.md (project); global rules in Warp Drive | ? | AGENTS.md | .roo/rules |
| Self-build (ask agent to extend)origin: Pi | – | – | – | – | core philosophy | – | – | – | – | – | – | – | – | – | – | – | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | ? | – |
| Output stylesorigin: Claude Code | native | /personality (none|friendly|pragmatic) | – | – | --system-prompt / SYSTEM.md | --system-prompt (replace main-session prompt) | – | – | – | – | – | – | – | – | – | – | ? | ? | ? | ? | --append-system-prompt / --append-system-prompt-file | ? | ? | ? | ? | --append-system-prompt | ? | – |
| Web search / fetchorigin: (generic) | web tools | limited | yes | /web | extension | yes | yes | yes | via MCP | yes | yes | @web | yes | via MCP | web fetch/search | Browser surface | yes (browser in VM; !deep for advanced search) | yes | #url context provider + MCP | Tavily web search (native tool; auto-triggered) + /extract URL fetch | native (WebSearch tool) | ? | Tavily MCP (built-in on Cloud) | yes (paste URL -> page content as context; no autonomous web search documented) | native agent web search (auto-triggered; Claude 4.x, GPT-5.x) | WebSearch / WebFetch | ? | yes |
| Image / multimodal inputorigin: (generic) | yes | yes | yes | images | yes | yes | yes | yes | yes (file picker attachments) | yes | yes | yes (+PDF) | yes | yes | yes | yes | yes (CLI Ctrl+V image paste; attachments) | yes | yes | yes (PNG/JPEG drag-and-drop or @ reference) | Ctrl+V (paste image attachment) | yes (attach image via paperclip) | yes (images via vision LLMs) | yes (screenshots, design drafts in chat) | image attachment in Agent Mode (Ctrl+G) | yes (drag-drop / paste / path) | yes (@ mention image attachment: .png/.jpg/.jpeg/.gif/.webp sent as native multimodal content to vision-capable models) | yes |
| Voice / dictationorigin: Aider (/voice) | – | – | – | /voice | – | – | – | – | – | Dictation | voice (Kilo) | – | – | – | voice input (hold space bar; Ctrl+X V) | ? | yes (inline voice recording button; voice follow-ups) | ? | ? | ? | ? | ? | – | ? | native voice input (microphone in Agent Mode) | ? | /voice (experimental; Ctrl+R to record) | voice (Kilo) |
| Browser / computer useorigin: Anthropic (Computer Use) / Cline | via MCP | – | – | – | extension | Computer Use (desktop automation) | – | yes | via MCP | Computer Use | browser | via MCP | live previews | – | Browser surface | Browser surface | yes, Chrome with Playwright CDP; visual QA + form-filling | Playwright (web app testing) | dev servers (background process management) | ? | Droid Control (launch apps, click, type, record) | ? | yes (BrowserToolSet / BrowsingAgent) | ? | Computer Use (sandboxed cloud environments, experimental) | ? | ? | browser |
| LSP integrationorigin: Zed / Crush | – | – | – | – | – | – | – | – | LSP-enhanced | – | – | yes | yes | native LSP | yes | yes | ? | ? | yes (LSP code intelligence: go-to-def, references, hover, diagnostics) | yes (JetBrains IDE ACP; semantic + syntax checks via IDE) | ? | ? | ? | LSP tool (listed among built-in Agent/subagent tools: Bash, Edit, Glob, Grep, Read, Skill, TodoWrite, WebFetch, WebSearch, Write, LSP) | built-in LSP (hover, go-to-def, diagnostics; Rust/Go/Python/TS/C++) | LSP tool | ? | – |
| Git commit / PR automationorigin: Aider (auto-commit) | commit cmds | yes | yes | auto-commit | extension | yes | yes | yes | yes | yes | yes | yes | yes | yes | /pr | yes | yes, ships PRs; Devin Review; auto-merge; PR resume | yes | yes (Web opens PRs; GitHub + GitLab integration) | yes (auto branch + commit + PR/MR via GitHub/GitLab actions) | GitHub Action (security --full: new branch + PR with security report) | yes (Remote Agents open PRs; Code Review bot) | yes (GitHub Action / fix-me label) | AI commit-message generation (one-click); no PR automation | native git dialog (Commit, Push, Publish, Create PR, View PR) | ? | yes (reads files, runs commands, writes code, and opens pull requests, approval-gated) | yes |
| Notebook (Jupyter) supportorigin: Cline | NotebookEdit | – | – | – | – | – | – | – | – | Jupyter | – | – | – | – | yes (agent mode) | ? | yes (hosts + populates Jupyter notebooks) | ? | ? | ? | ? | ? | Jupyter plugin (IPython cells) | ? | – | NotebookEdit | ? | – |
| Cloud handoff (local->cloud)origin: Windsurf | – | Codex cloud | – | – | – | – | – | – | – | Cline cloud | Kilo cloud | Background Agent | hand off to Devin | – | /delegate | Managed Agents | /handoff (CLI packages context + branch -> cloud session) | native (cloud-only) | Kiro Web (start local via IDE/CLI, run on cloud sandbox) | ? | Droid Computers (Bring Your Own Machine <-> cloud handoff) | ? | ? | ? | bidirectional local/cloud handoff (auto-handoff on sleep) | ? | & prefix (runs prompt in a Vibe Code Web sandbox) / /teleport (moves the whole session to Vibe Code Web) | Kilo cloud |
| Sandbox / permission controlsorigin: (generic) | permissions | sandbox | permissions | --yes | extension (gates) | auto-approval | yes | yes | permission queue; --yolo | auto-approve / YOLO | auto-approve | sandbox + net controls | Turbo (auto) | ? | permissions / auto-approve | Managed sandbox | yes, snapshot-isolated VM; /bypass mode (CLI); network/secret controls | cloud VM sandbox | Trusted Commands + Command Denylist; --trust-all-tools / --trust-tools (CLI headless) | brave mode (Off/Auto/On); allowlist.json; .aiignore | Autonomy levels (Off/Low/Med/High) + OS sandbox (--skip-permissions-unsafe) | Tool Permissions (allow/deny rules in settings.json) | Docker sandbox + approval flags | Sandbox (restricted, isolated execution environment for agent-generated commands; per-path read/write allowlist) + Auto-Run mode for MCP/commands | Agent Profile permissions (Confirm/Balanced/YOLO) + denylist | /permissions, /sandbox; -y / --dangerously-skip-permissions | agent approval modes (default/plan/accept-edits/auto-approve) + bash allow/deny lists + trusted-folder confirmation | auto-approve |
| Shared / team threadsorigin: Amp | – | – | /share | – | share session | channels | shared threads | – | – | – | – | – | – | channels (multiplayer) | – | Agent Manager | yes, Slack channel sharing; shared session links; #devin-runs | – | ? | ? | live session URL sharing (watch, comment, take over) | Share dialog (invite a teammate to view a Cosmos session; revocable) | ? | ? | Session Sharing (edit access: steer/approve/redirect) | ? | ? | – |
| Leaderboards / usage metricsorigin: Amp | – | – | – | – | – | – | leaderboards | – | – | – | – | team analytics | – | – | Copilot metrics API | ? | consumption analytics (daily/per-user/per-org via API v3) | ? | team usage analytics (Team plans) | ? | Factory Analytics (per-user usage breakdowns) | Analytics API (per-user activity, daily usage, per-editor/language breakdown, credit usage) | ? | TRAE Enterprise console: multidimensional usage analytics (per-client user/activity trends, code-generation rate, LLM call distribution, language distribution) plus rankings for MCP and enterprise agents | Enterprise Analytics API (per-user rollups, team metrics) | ? | ? | – |
| Org / policy controlsorigin: (vendor stacks) | enterprise | enterprise | – | – | – | – | team plans | – | – | – | – | team / SSO | enterprise SLAs | – | org policies / SSO | Enterprise (coming) | yes, RBAC, SSO/SCIM, IP allowlists, audit logs, enterprise guardrails | – | SAML/SCIM SSO via AWS IAM Identity Center; org management dashboard | Enterprise SSO; .aiignore enforcement; on-premises model hosting | SSO (SAML/OIDC) + SCIM directory sync | Team roles (Administrator / Member permissions) | org roles/permissions (Cloud/Enterprise) | TRAE Enterprise admin console (seat/member management, usage analytics, security policies) | Admin Panel (enforced org-wide AI/model/sharing/privacy policies; SSO) | ? | ? | – |
Showing 61 of 61 primitives · 28 harnesses · snapshot 2026-09-07. Cells show how each tool exposes the primitive; colour is the derived support level. “?” means unconfirmed this pass, not “no”.
Browse the matrix, methods, provenance, and CSV/JSON downloads →
The wider harness catalogue
The catalogue tracks 106 coding agents, from small terminal tools to IDEs and cloud platforms. The support matrix currently covers 28; the other 78 entries are candidates for future coding, not silent “no” values.
Claude Code
AnthropicGeneral agent
CLAUDE.md, hooks, subagents, skills, plugin marketplace; subagents can now spawn nested subagents up to depth 3 by default (v2.1.219, Jul 24 2026; was depth 1)
Codex
OpenAICoding agent
AGENTS.md and /goal; skills-based Spec Kit
opencode
Anomaly Innovations (formerly SST)Coding agent
Model-agnostic terminal agent; /share
Aider
Paul GauthierCoding agent
Repo map, diff edits, /architect-/ask, auto-commit, /voice
Pi
Mario Zechner (earendil-works)Minimal harness / toolkit
4-tool core (Read/Write/Edit/Bash); everything else is an extension
Qwen Code
Alibaba (Qwen)Coding agent
/plan, /goal, worktree, cron, rewind, fork subagent
Amp
Amp Inc. (spun out from Sourcegraph)Coding agent
Deep mode; Librarian/Oracle subagents; shared threads; leaderboards; Puck meta-agent (launched Jul 20 2026), a standing assistant on ampcode.com for launching, coordinating, and searching across other Amp agent threads
Goose
AAIF (donated by Block)Coding agent
MCP-first; recipes; simultaneous multi-model
Crush
CharmbraceletCoding agent
Charm TUI; LSP-enhanced; MCP-extensible
Cline
Open-source (Apache-2.0)IDE extension agent
Plan/Act split, Memory Bank, per-tool-call checkpoints, Workflows
Kilo
Anaconda (acq. 2026-07-15)IDE extension agent
Forked Cline via Roo; Boomerang/Orchestrator + custom modes. JetBrains surface rebuilt as a native Kotlin plugin supporting JetBrains split mode; multi-agent control room launched 2026-09-01: run multiple coding agents on different tasks at once, each in its own isolated git worktree, executed via existing Run Configurations, with diff and GitHub PR tracking from one IDE frame.
Cursor
Anysphere (acquired by SpaceX; all-stock ~$60B deal announced 2026-06-16, close confirmed by Cursor on 2026-08-14; Anysphere is now a wholly-owned SpaceX subsidiary)AI-first IDE + CLI
Pioneered AI-first editor; .cursor/rules; loads ~/.claude resources. Jun 2026: Compile conference (Origin git platform waitlisted, Composer 3 in training on Colossus). iOS app Jun 29, iPad app Jul 29. Router (Jul 22): intelligent per-request model routing. Cursor Start India plan (Jul 28). Cloud Agents Builds 3x faster (Aug 13).
Devin Desktop
CognitionAgentic IDE
Devin Local agent (Cascade sunset Jul 1 2026); ACP multi-agent orchestration; Devin handoff
Zed
Zed IndustriesAI-native editor (Rust)
Parallel Agents; channels; threads; native LSP; ACP; runs Claude Code
Copilot
GitHub / MicrosoftIDE + CLI agent
Custom agents (.agent.md), plan agent, Fleet mode, 8 hook types, GitHub MCP
Antigravity
GoogleAgent-first platform
Gemini CLI folds in; legacy Gemini CLI free/Pro cutoff Jun 18 2026
Devin
CognitionVendor platform
Autonomous AI software engineer; parallel cloud sessions; PR-native
Jules
GoogleCoding agent
Async cloud-VM agent; Gemini-powered; concurrent tasks; diff review
Kiro
Amazon / AWSAgentic IDE
Spec-driven dev; prompts to executable specs; property-based testing
Junie
JetBrainsIDE coding agent
Plans before coding; debugs with real debugger; #1 SWE-Rebench
Factory Droid
FactoryTerminal CLI agent
Agent-native SDLC; #1 Terminal-Bench (58.75%); model-agnostic
Augment Code
Augment CodeVendor platform
Context Engine; live whole-codebase index; persistent cross-session memory
OpenHands
All Hands AICoding agent
Autonomous agent edits code, runs commands, browses; formerly OpenDevin
Trae
ByteDanceAgentic IDE
AI-native IDE forked from VS Code with built-in agent
Warp
WarpTerminal CLI agent
Agentic terminal with multi-agent orchestration (Oz platform)
CodeBuddy
Tencent CloudVendor platform
Tencent's AI code editor; IDE, plugins, and CLI (CodeBuddy Code)
Mistral Vibe
MistralCoding agent
Mistral's coding agent / CLI with full-codebase understanding
Roo
Roo CodeIDE extension agent (defunct)
Forked Cline; Boomerang/Orchestrator + custom modes (archived May 2026)
Cosine (Genie)
Cosine AICoding agent
Proprietary code models; Swarm mode parallel child agents; every turn a reversible git commit
Refact.ai
SmallCloud (original creator; archived its own smallcloudai/refact and smallcloudai/refact-self-hosting GitHub repos on 2026-05-30, pointing users to the community-maintained fork JegernOUTT/refact for all new development)Coding agent
#1 open-source agent on SWE-bench Verified; local-first BYOK; integrates with GitHub, Docker, CI/CD
Snowflake CoCo
SnowflakeVendor platform
Data-native coding agent; understands Snowflake catalog, lineage, and RBAC; generates code grounded in real data permissions
Sweep
Sweep AIIDE extension agent
JetBrains-native; uses internal JetBrains static analysis tools; three modes: Ask, Agent (auto-edits codebase), Planning
Replit Agent
ReplitCloud coding agent
Full-stack cloud build agent; Agent 4 forks tasks into parallel branches then merges; sets up workspace, writes code, configures DB/auth/hosting autonomously
Zencoder
ZencoderIDE extension agent
Zenflow spec-driven multi-agent orchestration; RED/GREEN/VERIFY loops; one agent writes, second reviews, third audits for security and spec adherence
Baidu Comate
BaiduIDE coding agent
Zulu agent for end-to-end requirement-to-code generation; VS Code and full JetBrains suite integration
Qoder
Alibaba (Alibaba Cloud)Agentic IDE
Quest mode for multi-step autonomous task delegation; Knowledge Engine learns from codebase; supports up to 100k files and 26-hour agent sessions
Kimi Code
Moonshot AITerminal CLI agent
Built-in coder, explore, and plan subagents in isolated contexts; MCP integration; video input; ACP for IDE integration
MiMo Code
XiaomiTerminal CLI agent
Four-layer persistent memory (session, project, global, history) for long-horizon tasks; Compose mode; Dream background memory consolidation; 62% on SWE-Bench Pro and 73% on Terminal Bench 2 (vendor self-reported)
Huawei CodeArts
Huawei CloudAgentic IDE + CLI coding agent
Agent Team mode auto-assembles multi-agent dev team with concurrent execution; codebase indexing at tens-of-millions-of-lines scale; spec-driven development
Embedder
Embedder Tech Inc. (YC S25)Domain-specific coding agent (embedded firmware)
Autonomous build-flash-test-debug loop on real hardware; reads datasheets, writes firmware, flashes board, runs on-device tests, iterates autonomously
altimate-code
AltimateAIDomain-specific coding agent (data engineering)
OpenCode fork for data teams; 100+ deterministic SQL/dbt/lineage/FinOps tools; standalone terminal agent, embeddable under Claude Code/Codex
Zot
Patric Eckhart (independent)Terminal CLI agent
Single static Go binary; 4-tool core (read/write/edit/bash); session fork/branch/tree with .zotsession export; background subagent swarm; JSON-RPC extensions
ZCode
Z.ai (Zhipu AI)Agentic development environment
Long-horizon agent loop optimized for GLM-5.2; remote follow-up via WeChat/Feishu/Telegram
dbt Wizard CLI
dbt LabsDomain-specific coding agent (analytics engineering)
Native metadata engine indexing lineage, model health, tests, contracts, semantic definitions; builds/debugs/refactors/migrates dbt models
Codebuff
CodebuffAI (YC F24)Coding agent
Multi-agent TypeScript orchestration; codebase-aware agent coordination
Plandex
Plandex AICoding agent
Cumulative diff-review sandbox isolating AI changes across many files/steps until reviewed/applied, with rollback
Grok Build
SpaceXAI (rebranded from xAI, 2026-07-06)Coding agent
Extensible terminal coding agent on bespoke grok-build-0.1 model; interactive TUI, headless for scripts, or via API
Grok CLI
Superagent (YC-backed)Coding agent
Independent OSS terminal agent against public Grok API; sub-agents on by default; file edit, shell exec, web/X search, MCP
MiniMax Code
MiniMaxCoding agent
Code 2.0 (v3.0.48-49, Jul 13 2026): core session execution, state management and tool-calling rebuilt; official changelog cites cutting p95/p99 first-token latency by over 90% and reduced silent stalls/lost context; multiple independent Chinese outlets (Sohu, aibase, Sina/Readhub) report the rebuild is based on the open-source Pi Agent framework, though MiniMax's own English changelog does not name the framework
Crabcode
Blankeos (independent)Coding agent
Pure-Rust reimplementation of OpenCode UX (ratatui/crossterm TUI); dual PLAN/BUILD agent modes; multi-provider via models.dev
Trae Agent
ByteDance (SE Lab)Coding agent (open-source CLI)
General-purpose SWE agent CLI; file-editing, bash-execution, sequential-thinking tools; trajectory recording; multi-LLM support
Emergent
EmergentlabsCloud coding agent (vibe-coding platform)
Conversational multi-agent build loop: agents design, code, deploy full-stack apps from prompts; ownable, downloadable, GitHub-syncable code
Juggler
Julian Storer (independent, creator of JUCE/Tracktion/Cmajor)Coding agent
CRDT-backed (Yjs) tree-structured sessions instead of linear chat transcript; branching/backtrack/compare/edit threads; plugin-everything extension architecture (context items, strategies, commands as JS plugins); local Go/Wails webserver so desktop app, browser tab, and remote machine can all be simultaneous clients of one live session
Deep Agents Code
LangChainCoding agent
Sub-agents with isolated context windows for delegation; pluggable persistent memory across sessions (Memory-First Protocol); on-demand skills; built on the model-agnostic Deep Agents SDK (works with any LLM that supports tool calling)
Nanocoder
Nano Collective (independent community collective, not a company)Coding agent
Bring-your-own-model terminal agent (local Ollama or any OpenAI-compatible API/OpenRouter/Anthropic/Google); no telemetry, no paid tiers; governed and released by a volunteer open-source collective rather than a VC-backed vendor
CodeWhale
Hmbown (independent)Coding agent
Formerly DeepSeek TUI, renamed as the project moved beyond a single-model identity into a vendor-agnostic multi-provider runtime; Plan/Act/Operate mode split with Ask/Auto-Review/Full-Access permission tiers, OS-kernel sandboxing (Seatbelt on macOS, bubblewrap on Linux), and 'fleet' parallel workers that record every step to an append-only ledger for `fleet resume`
Aura
Ramen VRDomain-specific coding agent (game development)
Coding Agent that creates, edits, and fixes C++ and Blueprint files with deep understanding of engine source and project architecture, capable of self-correcting in real-time
Sarvam Code
Sarvam AICoding agent
Decomposes the coding workflow into three specialised agents, a planner, a worker and a verifier, to solve the common open-source-model failure mode of prematurely declaring a task complete; steerable long-running plan with resumable checkpoints; pay-for-completed-work billing where retries and dead ends are not charged; Sarvam's own Terminal-Bench 2.1 comparison claims about $2 per solved task versus $4.10-$27.80 for Claude Code and Codex
Grinta
Youssef Mejdi (independent)Coding agent
Local-first agent loop (control plane, execution, session history all run without cloud dependencies) built for long, failure-prone autonomous runs; real LSP and DAP integrations rather than shelling out to linters; durable event ledger with checkpoints and audit trails so a run can be inspected or resumed; provider-agnostic (OpenAI, Anthropic, Google, OpenRouter, Ollama, LM Studio)
Praana
amitkumardubey (independent)Coding agent
A deterministic compiler curates what the model sees each turn (tiered working memory, tool-output distillation, session checkpoints) instead of feeding the model the full transcript; optional cross-session Cognitive Memory extracts learnings at session end and surfaces a ranked digest on the next run; single Bun/TypeScript binary, provider-agnostic
Forall
Astrio LabsDomain-specific coding agent (formal verification / correctness proofs)
Generates spec-driven code alongside machine-checkable proofs of correctness rather than code alone; usable as a full CLI coding agent (install script, own account or OpenAI/OpenRouter key) or as a verify-only MCP server plugged into an existing agent (Cursor, Claude Code, Codex); currently covers TypeScript, Java and Rust
MandoCode
Armando Fernandez (DevMando, independent)Coding agent
.NET/Semantic Kernel harness built Ollama-first so it runs fully offline with no API keys by default (optional cloud model connection); MandoCode Desktop, a native Windows app (WinUI 3), shares the exact same underlying engine as the MandoCode CLI via a pinned git submodule so reasoning is identical across both surfaces, adding a multi-agent tabbed interface, integrated terminal, and git-aware file explorer on top
DevEco Code
Huawei / OpenHarmony SIG (AtomGit)Domain-specific coding agent (HarmonyOS development)
Terminal AI agent scoped to HarmonyOS/ArkTS app development (code writing, compile/build, run-on-device, doc lookup, runtime debugging, ArkTS issue fixing)
SolonCode CLI
OpenSolon / Noear open-source communityCoding agent
Built on the Solon AI framework in pure Java (Java 8-26 runtime support); pure-Chinese-prompt-driven agent design; approval-execution / auto-edit / read-only-plan work modes; persistent session history with rewind/redo; extensible via Skills, Agents, MCP, OpenAPI and LSP; one codebase ships identical agent logic across CLI, browser and a Tauri+React+Monaco desktop IDE
oh-my-pi
Can Boluk (independent), forked from Mario Zechner's PiTerminal CLI agent
Fork and rewrite of the already-tracked Pi harness into a coding-first surface with IDE-grade capability built in rather than a bare agent loop, including LSP integration, real debuggers (lldb, dlv, debugpy), persistent Python and Bun execution kernels, hash-anchored edits, 32 built-in tools and 40+ LLM providers
Peri
KonghaYao (independent)Terminal CLI agent
Built on the Agent Client Protocol (ACP) rather than a bespoke wire format, giving one 13MB Rust binary three access surfaces (terminal TUI, Zed IDE panel, headless CI); seven specialized sub-agents (coder, explorer, planner, reviewer, web researcher, verification, general-purpose) with pipeline and parallel execution; supports Claude Code plugins and skills as an input format without being a wrapper around the Claude Code binary
jcode
1jehuang (independent)Terminal CLI agent
Rust harness built around per-turn semantic-vector memory (cosine-similarity recall instead of replaying the full transcript), a swarm mode where multiple agent instances coordinate inside one repo with automatic conflict resolution, a custom terminal renderer (Handterm) for partial-line scrolling, and a self-modification loop where the agent can edit, build and reload its own source; supports 25+ model providers via unified OAuth/API
CleanSlate
TheWariend (independent developer)Coding agent
Fork of Visual Studio Code (MIT); Execution mode (full access) vs Plan mode (read-only analysis); can control a web browser to verify its own work; MCP server integration; BYOK across providers (OpenAI, Anthropic, Google, etc.) or managed models via optional CleanSlate Pro
DSCode
ThinkAny AI (GitHub org thinkany-ai)Coding agent
DeepSeek-native runtime defaulting to DeepSeek V4-Flash, with multi-provider support (OpenAI, Anthropic, others); local-first sandboxed tool execution; Git-based patching; parallel agent work; readable/transparent token-cost reporting
Shelley
Bold Software (maker of exe.dev)Coding agent
Successor to Bold Software's earlier Sketch coding agent; multi-conversation, multi-modal, multi-model, single-user agent built for persistent VMs (Sketch assumed containers only lasted one conversation); ships with no built-in authorization or sandboxing by design, bring-your-own
Deep Code
Vegamo / lessweb (independent, community)Coding agent
Deep-thinking mode, reasoning-effort control, Agent Skills (~/.agents/skills and ./.deepcode/skills), MCP integration, optimized for the DeepSeek-V4 model
Reasonix
esengine (independent, community)Coding agent
Go kernel distributed as a single static binary; config- and plugin-driven; engineered around DeepSeek's prefix-cache stability to keep token costs low across long-running sessions; optional dual executor+planner model setup
Neal
navels (independent)Coding agent
Plan-driven multi-agent coding loop (`neal run/plan/execute`) with separate planner, coder, and reviewer roles, each on a provider/model the user picks; coder and reviewer are deliberately run on different vendors so review is adversarial rather than one model grading its own work; each scope starts the coder from a fresh context to prevent drift across long tasks; reviewer access is enforced read-only across provider adapters
Muse Code
MetaTerminal CLI agent
Persistent async background sub-agents that stay active for the whole session; parallel sub-agents isolated in separate Git worktrees; local event log for every model call/tool run/edit for replay-exact, restart-safe crash recovery; powered by Muse Spark 1.2
Prime Agent
Prime IntellectTerminal CLI agent
Recursive Language Model (RLM) runtime + Continual Harness: context-as-a-variable, programmatic tool/sub-agent calling via a persistent IPython kernel, multi-agent messaging, self-modifiable harness state
Agent AFK
Griffin Long (independent)Terminal CLI agent
Harness built for unattended, away-from-keyboard runs across four surfaces sharing one session manager (one-shot CLI, REPL, cron-friendly headless daemon, Telegram bot); editable prompts/gates/routing/skills/traces as code
agentty
1ay1 (independent)Terminal CLI agent
Native C++26 implementation; single ~13.6MB static binary, sub-ms cold start, no runtime/npm install needed; lives at the bottom of the terminal preserving scrollback rather than taking over the screen
Waveloom
Menfre01 (independent)Terminal CLI agent
Go binary built on Charm's Bubble Tea/Glamour/Lip Gloss; engineered around DeepSeek V4 prefix-cache economics; four-tier context compaction; Fork/Cold/Explore sub-agent modes; post-edit LSP diagnostics for Go/Rust/TS/C/C++
AtomCode
AtomGit (OpenAtom Foundation open-source community)Terminal CLI agent
Rust-built terminal coding agent hosted on AtomGit (China's OpenAtom Foundation open-source platform); positions itself as an open-source alternative to Claude Code / Cursor Agent, connects to any OpenAI-compatible model (DeepSeek, Qwen, etc.), supports macOS/Linux/HarmonyOS PC/Windows
CodeFlicker
Kuaishou (StreamLake)Agentic IDE
Agentic coding platform from Kuaishou's StreamLake unit; web-based IDE with intelligent code completion, real-time preview/debugging, and repo-aware agentic planning. Offers Kuaishou's own Kuaiyi LLM 3.0 plus Claude Sonnet 4/4.5 and Gemini 2.5 Pro as selectable models
KlaatCode
KlaatAITerminal CLI agent
Open-source terminal coding agent with its own read/edit/run/verify tool loop, built on Anthropic's Claude Agent SDK; a thin terminal client to KlaatAI's hosted routing service (Klaatu), which picks the model per task to cut cost rather than the CLI wrapping the Claude Code binary; headless/CI mode and API server mode also ship
bbarit-oss
BBARITTerminal CLI agent
Open-sourced standalone version of the terminal-native coding agent originally built into BBARIT Terminal (vendor site names the product 'bbarit-oss'; GitHub repo is bbarit/bbarit-agent-oss); single static Rust binary, 15+ LLM providers/1,000+ models, own assistant-tool-call-result loop, full-screen TUI plus one-shot/JSON/headless modes
Rovo Dev
AtlassianVendor platform
Atlassian's agentic coding platform; ships a CLI as an Atlassian CLI (ACLI) extension plus IDE surfaces, with deep Jira/Confluence integration; own agent loop, not a wrapper on another vendor's binary
GitLab Duo CLI
GitLabTerminal CLI agent
Terminal agent (glab duo cli) bringing GitLab Duo Agentic Chat to the terminal, with project/pipeline/agent-config awareness; understands codebase structure and cross-file functionality and performs multi-step development tasks autonomously; runs GitLab's own agentic capability rather than launching an external agent like Claude Code or Codex CLI
Kiro Crew
Amazon / AWSCoding agent
Persistent multi-agent orchestration workspace built around the Kiro CLI via the Agent Client Protocol (ACP); schedules recurring work, retains cross-session memory, and coordinates multiple agents/sub-agents across sessions without a developer at the keyboard
Ante
Antigma LabsCoding agent
Self-contained ~15MB Rust agent harness with a managed llama.cpp inference engine that auto-detects local GGUF models for fully offline use, alongside support for a dozen-plus cloud LLM providers; ships multi-agent, skills, and MCP support with no runtime dependencies
Code Puppy
mpfaffenberger (independent GitHub maintainer)Coding agent
Open-source terminal coding agent that reads, creates, edits, and deletes files and runs shell commands, with multi-model support across 65+ providers, JSON/Python-configurable custom agents, and DBOS-backed durable execution for crash recovery
Bough
andreylukin (independent GitHub maintainer)Coding agent
Coding agent that responds to each turn by generating and executing a single JavaScript program (with real loops, branching, and composition) against the actual repo checkout via nineteen host functions, rather than issuing sequential tool calls; uses numbered-line view/patch edits to avoid string-escaping issues and detect stale-edit conflicts
DeepSeek Harness
DeepSeek AITerminal CLI agent
Plugin-based agent harness (dsh CLI) where models, tools, sandboxes, session storage, UI and the agent loop itself are all plugins on the Cordis framework
Hax
Oleksandr Chekhovskyi (independent)Terminal CLI agent
Minimalist terminal-native coding agent written entirely in C; single native binary with a small dependency set, starts instantly, uses only a few MB of memory; own tool-use/agent loop (Read/Write/Bash-style tools, streaming Markdown output, session resumption, transcript view) rather than shelling out to Claude Code/Codex; supports OpenAI, Anthropic, Codex-via-ChatGPT-subscription, OpenRouter, llama.cpp and Ollama as model providers
Bullet
Bullet (Y Combinator S26)Coding agent
YC S26 coding agent built by Yale CS grads (ex-AppLovin/Citadel); own router, tools, and agent loop shared between a CLI (npm @trybullet/cli) and a Mac desktop app, positioned around latency: auto-picks model/reasoning tier per prompt, parallelizes searches/reads/commands, and uses targeted code search instead of full-repo embedding
Genie Code
DatabricksDomain-specific coding agent (data engineering / data science)
Unity Catalog-grounded context selection; plans multistep data engineering/ML/BI workflows for approval before autonomous execution; writes and debugs SQL and Python pipelines
Hoplite
Hoplite (YC S26)Cloud build agent
In-house coding harness built rather than wrapping Claude Code/Codex, so the team can ship agent features without waiting on Anthropic/OpenAI; tasks dispatch from Slack, Linear, iMessage, or automatically from Sentry issues into cloud sandboxes
Blacksmith codesmith
BlacksmithCloud build agent
Cloud coding agent embedded in Blacksmith's CI/validation loop: delegate feature/bugfix tasks, and it also diagnoses and autofixes CI failures to keep PRs green, distinct from a general-purpose coding CLI
Gajae Code
Yeachan-Heo (independent maintainer)Coding agent
Provider-agnostic external harness (clarify before guessing, plan before mutation, track every goal to completion with evidence) that runs standalone from any repo/worktree and layers onto 50+ existing coding-agent subscriptions and model providers (Claude, Codex, Cursor, Copilot, OpenCode, Kimi, GLM, MiniMax, Grok, Qwen, local runtimes) instead of billing separately
Neovate Code
Ant Group / Alipay (Alipay Experience Technology Department)Coding agent
Instruction-driven generate/fix/review/test workflows with interactive and headless modes; AGENTS.md rule files; conversation continuation/resumption; provider-agnostic (OpenAI, Anthropic, Google); monorepo also ships a VS Code extension
Metis
Wholiver (independent)Coding agent
Recursive L0-L4 multi-agent role split (coordinator, planner, implementer, reviewer, verifier); SQLite-backed persistent memory; Git worktree isolation per task; model-agnostic (OpenAI, Anthropic, DeepSeek, Gemini, Groq, Ollama, vLLM, or OrcaRouter)
Devx
ApvCode (independent)Coding agent
Dual PLAN (read-only architect) / AGENT (autonomous file edits, shell exec, package install) mode split; snapshot-based /undo rollback; multimodal vision via clipboard image attachments; built-in TypeScript/JavaScript/Python/Rust compile diagnostics; session history with unique session IDs
TheGitAI
Atlast TechnologiesCoding agent
Single-session investigate/build/verify/document/recover agent loop; checkpointed, journal-backed edit recovery; curated frontier model selection (site shows GPT-5.6 Sol) rather than open BYOK across arbitrary providers
VT Code
vinhnx (independent maintainer)Coding agent
Rust terminal agent with safe file operations, ripgrep search, ast-grep symbol maps, fuzzy discovery, and project indexing; multi-layered security model (tree-sitter-bash command validation, execution policy, sandbox isolation); 21+ built-in LLM providers plus local inference via Ollama, LM Studio, and llama.cpp
Aurict
Aurict (GitHub org handle; underlying company/individual identity unconfirmed)Coding agent
Nine specialist sub-agents (explore, code, review, test, debug, docs, security, performance, analytics) running as isolated workers; automatic codebase/framework context detection on startup; graduated shell-command approval (passive / active-lite / full security modes) with audit trails; multi-provider BYOK (Anthropic, OpenAI, Google, xAI, Azure, AWS Bedrock, Ollama, OpenRouter)
Kit
SpeakeasyCoding agent
Gives the model a single composable tool for writing and running programs rather than exposing many discrete tools; ships as one static binary with kit tui (interactive terminal client), kit prompt (one-shot task), kit acp (editor protocol server), and kit serve (headless daemon) entrypoints
Jern
Jern (vendor name as shown on jern.ai; no separate parent company disclosed)Cloud build agent
Policy pinned by digest for the whole session, isolated execution machine, credentials kept out of the agent process, hard token caps metered by the user's own provider, encrypted evidence storage; delivers all work exclusively as a pull request on an isolated branch (or a labelled failure), never auto-merges; the open-source Apache-2.0 runtime is described as the same agent that runs in the hosted cloud, runnable on your own laptop
Letta Code
Letta (MemGPT creators; UC Berkeley Sky Computing Lab spinout, founded by Charles Packer and Sarah Wooders)Coding agent
Memory-first coding agent built on Letta's own stateful-agents SDK; agents persist learned codebase knowledge across sessions instead of starting fresh each run; Letta's own blog post (self-reported, not independently verified against the public Terminal-Bench leaderboard) claims Letta Code is the number-one model-agnostic open-source coding harness on Terminal-Bench, a narrower category-specific claim distinct from Factory Droid's separately tracked overall '#1 Terminal-Bench (58.75%)' figure
OpenBitFun
GCWing (GitHub org handle; company identity unconfirmed)Coding agent
Rust agent runtime with a Tauri desktop UI; a built-in Code Agent plans, edits, tests and commits inside real git repos with filesystem, terminal, git, browser and desktop-app access, alongside general-purpose Mini Apps and Computer Use
CrewCoder
OnPoint-Dev-Tools (GitHub org)Coding agent
Evented agent loop emitting typed JSON events (assistant deltas, tool calls, approvals, usage) over stdout or SSE; transactional file read/write/edit, bash and background-job execution, LSP-backed symbol editing, with size-limited and redacted tool output
Primitive lineage
These entries record where a practice was popularised in coding-agent products. “Originated by” is an attribution within this product category, not a patent or priority claim.
| Primitive / concept | Originated by | Context |
|---|---|---|
| Plan mode / Plan-Act split | Cline | Introduced Plan Mode against the immediate-execution trend; 'frontload context, then act' |
| Boomerang / Orchestrator | Roo Code | Delegates subtasks to specialist modes; only the summary returns to the parent |
| Checkpoints after each tool call | Cline | Auditable shadow-Git checkpoints per tool call, not just per prompt |
| Memory Bank | Cline | Persistent project memory as plain markdown in the repo |
| AGENTS.md | OpenAI Codex | Now a Linux Foundation open standard across 60k+ repos (OpenAI, Google, Cursor, Factory, Amp) |
| /goal (persistent objective) | OpenAI Codex | First native persistent-objective primitive; later mirrored by Qwen Code |
| CLAUDE.md, /init, hooks, subagents, skills, plugin marketplace | Claude Code (Anthropic) | Anthropic's agent-architecture primitives |
| MCP (Model Context Protocol) | Anthropic | Created by Anthropic; adopted by OpenAI Codex, Goose, Crush, and others |
| Repo map, diff edits, /architect-/ask, auto-commit, /voice | Aider | Early terminal coding-agent conventions |
| Custom modes (role personas) | Roo Code | Code/Architect/Ask/Debug/Orchestrator + community Mode Gallery |
| Minimal-core / self-build harness | Pi | 4-tool core; plan mode, subagents, MCP, permissions are all user extensions |
| Steering / follow-up mid-run | Pi | Enter steers a running agent after current tool call; Alt+Enter queues a follow-up |
| Scheduled / cron agent tasks | Qwen Code | Cron-scheduled autonomous coding runs |
| AI-first editor; .cursorrules | Cursor | Pioneered the AI-first editor category and rules files |
| Cascade agent + Workflows + workspace memory + intent tracking | Windsurf | Autonomous multi-file agent; reusable named Workflows; tracks edits/terminal/clipboard |
| Deep mode (autonomous research) + shared threads + leaderboards | Amp | Extended-reasoning research; team-oriented thread sharing and metrics |
| Parallel Agents + channels + native-LSP editor | Zed | Rust editor; multiplayer channels; agent client protocol (ACP) |
| Fleet mode / plan agent / .agent.md / 8 hook types | Copilot | Parallel subagent orchestration + GA custom/plan agents |
| Computer use / browser control | Anthropic (+ Cline) | Anthropic Computer Use; Cline shipped in-IDE Computer Use |
| Spec-driven specify->plan->tasks->implement | GitHub Spec Kit | Cross-tool methodology; SPARC was a precursor in the Roo/ruv ecosystem |
Dynamic workflows: three flavours
“Workflow” is not a consistent feature name. The matrix separates three mechanisms that tools often group under it.
Scripted / declarative
A named sequence defined once and invoked again. Examples include Cline Workflows, Windsurf Workflows, GitHub Spec Kit, and SuperClaude /sc:workflow.
Orchestrated / delegated
A parent routes subtasks to specialist agents and receives compact results. Examples include Roo Orchestrator, Copilot Fleet, Amp subagents, Antigravity Agent Manager, and Pi team extensions.
Autonomous loop
The harness continues through bounded steps without a new prompt for each one. Examples include Cline Focus Chain, Qwen autonomous goals, Windsurf Turbo, Amp Deep mode, and extension-based loops.
Extensions: the portability layer
Some capabilities arrive through reusable packs rather than the harness itself. GitHub Spec Kit has the broadest cross-tool coverage in this dataset, with support for more than 24 agents.
GitHub Spec Kit
Cross-tool CLI (specify)Spec, constitution, plan, tasks, implement
Works with: 24+ agents incl. Claude Code, Codex, opencode, Cursor, Copilot, Antigravity, Pi, Qwen, Goose
Most portable; /speckit.* (Codex uses $speckit-* skills)
SuperClaude
Frameworkplan/design, implement, brainstorm, task/spawn, workflow, estimate, spec-panel, review, index, research
Works with: Claude Code
~30 /sc: commands + personas
Superpowers
PluginSpec-driven plan -> implement loop
Works with: Claude Code
One-command install; SessionStart hook
BMAD-METHOD
FrameworkMulti-role spec, task breakdown, brainstorming, adversarial review
Works with: Claude Code
Best for multi-human-role teams
oh-my-opencode
Plugin pack/init-deep, /ralph-loop (autonomous), /refactor, /start-work
Works with: opencode
Adds workflow commands core opencode lacks
SPECLAN
Plugin (cc-marketplace)Goal/Feature/Requirement hierarchy; plan -> implement -> review
Works with: Claude Code
Goal-as-artifact w/ lifecycle states
claude-flow
OrchestratorParallel agent swarms, orchestration, RAG
Works with: Claude Code + others
Swarm intelligence
agent-team orchestrators
CLI orchestratorsGoals, todo->review->done state machine, auto-retry, parallelism
Works with: Claude Code, Codex, Cursor
Typed agent teams + TUI
awesome-pi-agent / pi packages
Package ecosystemplan mode, subagents (Fleets/Crews), permission gates, MCP, checkpointing, sandboxing, themes
Works with: Pi
50+ example extensions to fork
Pi Commander
Companion appKanban dashboard, spec wizard, Jira integration
Works with: Pi
GUI layer over Pi multi-agent
pr-review-toolkit
Plugin (Anthropic)Specialized PR review agents
Works with: Claude Code
Official Claude Code plugin
mcp-memory-service
MCP server + hooksPersistent memory, context injection on session start
Works with: Claude Code (+MCP agents)
Memory awareness via hooks
Morph Fast Apply
MCP serverHigh-speed diff/edit apply
Works with: Claude Code, Codex, Cursor, any MCP agent
~10.5k tokens/sec apply
Vercel Skills.sh
Skill ecosystemReusable agent commands (skills)
Works with: MCP / agent CLIs
Open ecosystem
Special case: Pi
Pi deliberately ships a four-tool core: Read, Write, Edit, and Bash. Planning, subagents, MCP, permissions, and checkpoints are installed or built as packages. Its many “extension” cells therefore describe the architecture rather than missing functionality.
Native features include steering and queued follow-ups, session rewind, branching and sharing, AGENTS.md, portable Skills, and headless output.
Caveats
- Coverage. The matrix contains 28 of 106 catalogued harnesses; catalogue membership does not imply matrix coverage.
- Unconfirmed cells. 228 of 1,708 cells are marked “?”. This means unconfirmed, not unsupported.
- Fast drift. Coding agents move capabilities between commands, modes, and extensions as their interfaces change.
- Interface classes. A named mode and a typed slash command both count as built-in support, but the cell preserves the distinction.
- Claude Code /goal remains unconfirmed because third-party guides mention it while Anthropic’s official documentation does not.