Hermosa Research·Synthesis
Harness-as-a-Service: Agent Runtimes, SDKs, and Sandboxes
A comparison of 46 hosted runtimes, agent SDKs, code sandboxes, and browser services. The article draws from a source-backed catalogue of 109 AI infrastructure products across 17 categories.
- Published
- 21 June 2026
- Data as of
- 6 September 2026
- Research type
- [synthesis]Distils many external sources into a position. The rigor lives in source verification and per-claim attribution.[tentative]Reasonably supported, but parts are still unverified or liable to drift.
- Topics
- [agents][cloud][infrastructure]
Method: We classified 41 harness-layer products from a broader 95-service catalogue using official product and documentation pages. Vendor descriptions are not independent performance tests.
Background
Agent infrastructure now spans four separate layers. Hosted runtimes operate the agent loop, SDKs let teams own it, sandboxes isolate code, and browser services provide managed web sessions. Treating them as one product category hides the operational choices.
This comparison covers 46 products across those layers. It is a companion to AI Coding-Agent Primitives, which compares locally used coding-agent interfaces rather than the infrastructure behind deployed agents.
Key findings
- 1The harness layer is four distinct markets. 22 hosted runtimes operate agent loops, 10 SDKs expose the loop to developers, 10 sandboxes isolate code, and 4 browser services provide managed web sessions.
- 2Durable execution separates hosted runtimes. Long-running agents need explicit checkpoint, retry, resume, and scheduling behaviour. Inngest, Trigger.dev, Convex, LangGraph Cloud, and Vercel Eve expose different models for that work.
- 3Sandbox products make different isolation choices. The catalogue includes Firecracker microVMs, gVisor containers, pre-warmed pools, and general-purpose machines. Runtime, startup, session length, filesystem policy, and network policy matter more than a single winner label.
- 4MCP reduces tool coupling, not platform coupling. Model Context Protocol appears across hosted runtimes and SDKs, making tools easier to move. State, execution, identity, and observability can still bind an application to its runtime.
- 5State is the architectural decision. Durable actors, isolated processes, thread stores, and reactive databases make different promises about replay, concurrency, persistence, and recovery. Those promises shape the application more than the model API does.
- 6Pricing units do not compare cleanly. The products charge in credits, tokens, invocations, compute time, reviews, or browser hours. Cost comparisons need a defined workload, including idle time, retries, and retained state.
Anatomy of a harness
A production agent combines several functions: the loop, model access, tools, execution, state, and observability. Products cover different subsets, so two services described as “agent platforms” may leave teams with very different responsibilities.
The harness as a system
The agent loop at the centre; everything else is what it depends on at runtime. Each block is a layer of the AI Infrastructure stack.
The practical comparison is the vendor boundary. The diagrams below show three common ways to divide ownership between a platform and the team building on it.
Three ways to fill the same picture
Same six blocks, same arrows. What changes is where the vendor draws its boundary, and which blocks you bring yourself.
All-in-one reduces integration work but concentrates switching costs. Hybrid keeps a hosted core while bringing selected layers, often models or memory, under team control. À la carte maximises choice and operational responsibility.
The catalogue
The 46 entries below are grouped by layer. Each record covers hosting, models, execution, state, tools, browser and memory support, human approval, observability, pricing, maturity, and source. The data page adds the wider inference, memory, evaluation, media, and safety stack.
service 1 of 46 · Hosted runtime
AgentZ
launched 2026-08-27
both·free plan available; paid/enterprise tier pricing not publicly detailed in the launch announcement·multi-model (OpenAI, Claude, Grok, and others; bring-your-own-LLM)·workspace hierarchy: Organizations, Workspaces, Agents, Workflows, and Sandboxes, with users/roles applied across that structure·tool-level permissions plus runtime credential injection, reusable Skills, and workflow triggers·unknown·unknown·unknown·per-agent zero-trust sandbox with filesystem, domain-whitelisting, and network-access controls (backend not named by vendor)·workflow-based execution with triggers (scheduled or event-driven runs)·audit traces across agents and workflows
Summary
A model-agnostic enterprise platform to build, run, and govern AI agents, bundling per-agent zero-trust sandboxes, workflows, role-based access, runtime credential injection, and audit traces, deployable as SaaS, on-prem, or air-gapped.
Notes
New enterprise agent-management vendor; distinct from tracked hosted-runtime platforms (AgentCore, Foundry Agent Service, Gemini Enterprise Agent Platform) in bundling zero-trust sandboxing, credential injection, and air-gapped deployment as its primary sell rather than starting from a coding/browsing agent product. Product page confirms the 'Zero Trust Sandbox Platform' framing; skills/MCP/workflow claims are per the GlobeNewswire launch release.
Browse all 109 infrastructure products and CSV/JSON downloads →
Hosted runtime
The vendor operates the agent loop and its execution lifecycle. This reduces operational work but couples state, tools, and recovery behaviour to the service.
Examples: Anthropic Claude Agent SDK, Cursor Cloud Agents, Vercel v0+Eve, Cloudflare Agents, Microsoft Foundry, Inngest AgentKit, Trigger.dev, Convex Agents. (22 tracked).
SDK / framework
A library for application-owned orchestration. Teams control the loop and deployment, then assume responsibility for state, scaling, and recovery.
Examples: OpenAI Agents SDK, CrewAI, Mastra, Pydantic AI, smolagents, AWS Strands, Microsoft Agent Framework, Pi (pi-mono). (10 tracked).
Sandbox
An isolated environment for agent code, shells, and file operations. Compare isolation, startup time, session duration, filesystem policy, and network controls.
Examples: E2B, Modal, Daytona, Fly Machines / Sprites, Vercel Sandbox. (10 tracked).
Browser tool
A managed browser session exposed through an API. Persistence, proxies, browser control, and anti-bot handling vary, and vendor claims are not independently tested here.
Examples: Browserbase, Steel, HyperBrowser, Apify Actors. (4 tracked).
Choosing a layer
Use a hosted runtime when managed execution and recovery matter more than portability. Use an SDK or framework when the team needs to own the loop and deployment. Add a sandbox for isolated code execution, and a browser service only when the runtime does not already provide the required web access.
For any agent that survives more than one request, document the state model before choosing the product. Threads, durable actors, sessions, and checkpoints have different replay, concurrency, and recovery guarantees.