Hermosa AI
Loading...
Topics
Multi-Agent Interaction Patterns
45 patterns
[
fortnightly
]
2026-07-06
Build-Verify-Reflect
A builder produces output, a verifier (often deterministic code) checks it, and on failure a reflector feeds the error back for a bounded retry.
Standard Operating Procedures (SOP)
A complex task is decomposed into ordered phases with role-bound agents and gated hand-offs, an assembly line rather than a free-for-all conversation.
Role Specialization
Agents are given distinct personas, expertise, and objectives so the team probes a problem from genuinely different angles.
LLM-as-Judge
A separate strong model scores candidate outputs against a rubric, routing to approve / rework / reject.
Arena Rating (ELO)
Pairwise battles between agents accumulate into persistent ratings (ELO, or a match-order-invariant MLE) that rank them over many rounds.
Independent Sampling
Several agents work the same task in isolation, their answers sealed from one another until a synchronised reveal or vote.
Structured Debate
Two agents argue opposing positions across structured rounds while a judge scores argument quality, not just the conclusion.
Mixture-of-Agents
Layered ensembling: each layer of agents reads the previous layer’s outputs as auxiliary context, and an aggregator synthesises a final answer.
Consensus Triangulation
Several models independently analyse the same input, then debate; the output is a consensus map of agreement, divergence, and reasoning fault lines.
Negotiation & Equilibrium
Agents with explicitly competing utility functions negotiate over shared state; the outcome emerges from the interaction, not from either side alone.
Adversarial Stress Test
One agent builds something robust; an adversary with opposing objectives probes it for weaknesses; an arbiter scores severity.
Competitive Arena
Agents know their standing and their opponent’s, and that competitive context, injected into the prompt, measurably shapes strategy and draws out deeper capability.
Evolutionary Selection
A population of candidate programs is produced by an ensemble of differently configured LLM samplers in parallel; a separate evaluator scores them; the fittest survive and are mutated into the next generation, gradient-free search in solution space.
Teacher-Student Transfer
A stronger “teacher” demonstrates and critiques while a cheaper “student” attempts the task and improves over rounds, distillation at inference time.
Dynamic Topology Routing
Each round, every agent emits a need / offer descriptor; a semantic matcher rebuilds a sparse directed communication graph for that round only.
Topology-as-Task-Routing
Classify a task’s decomposition into a canonical shape (parallel / sequential / hierarchical / hybrid) and pick the orchestration topology per query.
Group-of-Agents
Model collaborative groups as atomic units and wire inter-group connections, instead of node-centric edges, explicit divide-and-conquer structure.
Two-Timescale Co-evolution
Co-evolve agent capability fast and team topology slowly, a meta-controller does agent birth/death and edge edits every few rounds under an edit budget.
Concurrent Shared State (Notify-and-Repair)
Let agents write a shared artifact concurrently; on conflict, notify affected agents and have them selectively repair, rather than locking or aborting.
Spawn-and-Isolate Fan-Out
An orchestrator owns the full context and spawns context-isolated workers that return one compressed artifact each; a dedicated aggregator merges them.
Adaptive Mixture-of-Agents
The layered experts → aggregator ensemble, made cheap by skipping the aggregator entirely when the experts already agree (37–90% of calls).
Blackboard Capability Volunteering
A central agent posts requests to a shared blackboard; subordinate agents self-select and volunteer responses based on their own knowledge and competencies, without a coordinator needing to know each agent's expertise.
Market Belief Trading
Agents act as market participants, updating and trading probabilistic beliefs with each other; price signals aggregate local epistemic states into a shared, calibrated, truth-seeking outcome without a judge or external enforcer.
Latent Working Memory
Agents bypass the token layer entirely, sharing last-layer hidden embeddings through a pooled latent working memory so that information exchange is lossless and sub-symbolic rather than serialised into text.
Transactive Trajectory Pool
Producer agents deposit completed task trajectories into a shared, indexed repository; consumer agents retrieve those trajectories as procedural context for new tasks, with no live coordination and no requirement that producers and consumers ever co-exist.
Reasoning-Tree Divergence Audit
Multiple agents produce independent reasoning traces that are atomised into steps and merged into a shared semantic tree; an adjudicator resolves conflicts by comparing competing branches only at their critical divergence points rather than re-reading full traces.
Peer Elicitation Game
A generator produces outputs that are evaluated by a panel of discriminator agents from distinct base models; discriminator utilities are computed via a mutual-information score that provably incentivises truthful reporting without requiring ground-truth labels, and the system converges to a truthful Nash equilibrium.
Canonical Merge (CvRDT Cache Fusion)
Multiple agents' latent reasoning traces (KV-caches) are fused via a content-addressed, order-independent merge: fragments are ordered deterministically by content (mean K-norm at a middle layer) so the union of any subset of agents' latent state renders to a byte-identical result regardless of arrival order.
Conditional Escalation
An executor agent handles tasks autonomously by default; on failure or low confidence it escalates to a stronger advisor, which counsels but never takes over execution.
Progressive Autonomy (Trust Ledger)
The system tracks a mutable trust score that shifts between supervision modes (watch, queue, auto), earned through consistent correct behaviour over time rather than granted upfront.
Failure Composting
Operational waste (failed runs, error logs, rejected outputs, edge-case discoveries) is systematically collected and transmuted into durable system improvements: new rules, updated gates, or refined prompts, with a human signature required before any rule takes effect.
Localized Multi-Agent Debate
Agents represent their reasoning traces as typed nodes, locate the earliest point where their claims conflict, and restrict debate to only that local segment instead of re-exchanging full traces. A guarded resolution step extends a shared committed state so a resolved segment cannot be reopened by later conflicts.
Causality-Inspired Semantic Packet Scheduling
Inter-agent messages are decomposed into semantic packets (evidence, requests, and similar units), each packet's predicted contribution to a downstream decision is estimated via causality-inspired replay valuation, and only packets whose predicted value exceeds their transmission cost are forwarded, subject to bandwidth, latency, deadline, and receiver-context constraints.
Persona-Matched Partner Selection
A host agent recruits a teammate at runtime from a pool of candidate agents that differ only in persona description, with capability held equal. In a controlled study across 375 trials, hosts selected partners by task-type stereotype (an open persona for creative work, a conscientious persona for most other categories) and favored complementary personas over self-similar ones, the opposite of human homophily.
Dynamic Debate Relationship Regulation (DEAR)
Mutual referencing among debating agents is modelled as an explicit, mutable graph of Debate Relationships. A Selection RL-Agent dynamically rewires which peers each agent attends to each round based on measured group-level consensus/divergence evidence, and a Behavior RL-Agent adapts each agent's generation strategy in response, the two optimised jointly as a sequential multi-agent RL problem.
Imprecise Belief Fusion
Agents represent beliefs as formulas in a propositional language and combine them pairwise with peers via a parametrised fusion operator; tuning the operator toward imprecision produces a less committal fused belief whenever two agents' beliefs diverge, rather than forcing immediate convergence to one side.
Reputation-Gated Cooperation
Agents play a repeated donation game where each decides how much to give a partner based on that partner's observed history of past donations; strategies are written as evolving natural-language prompts, and only strict exclusion of past defectors, not reward of past cooperators, predicts resistance to free-rider invasion.
Contradiction-Conditioned Replanning
A hierarchical LLM planner pairs with a Bayesian module that tracks a teammate's evolving strategy; instead of replanning on every new observation or ignoring a strategy shift entirely, the agent interrupts its current skill only when the partner's actions directly contradict the currently inferred skill.
Open-Strategy Transparency
Each agent's decision strategy is written as a natural-language document that every other participant can read before an interaction; an LLM adjudicator interprets one agent's decision in light of the other's published strategy, so conditionally cooperative strategies can be verified from the text itself rather than inferred from past behaviour.
Federated Memory Reconciliation (MELD)
Distributed agents each keep a sovereign memory modelled as a knowledge graph; incoming claims from peers are admitted through a five-outcome procedure (insert, merge, relate, conflict, or reject) using claim-key identity, embedding similarity, and natural-language-inference verdicts, then applied through a single auditable Patch, with detected contradictions preserved for later adjudication instead of being silently overwritten.
Budgeted Audit-Point Placement
An auditor agent is scheduled across a running multi-agent workflow modelled as a dynamic dependency graph: under a fixed audit-call budget, audit points are placed greedily on influential and long-unaudited regions to minimise cumulative unchecked exposure, and each trusted audit point becomes a recovery anchor so failures roll back locally instead of to the start.
Reward-Guided Topology Generation
The multi-agent communication graph is emitted by a trained autoregressive generator, and that generator is then fine-tuned RLHF-style against a learned reward model that jointly scores task correctness and structural compactness, so the topology designer itself is optimised to produce sparse graphs that keep accuracy while cutting token cost.
Epistemological Friction Loop
A deliberately high-entropy generating agent produces speculative, hallucination-tolerant output; a web-grounded evaluating agent tests it against external evidence; and a low-entropy semantic bottleneck between them filters noise and repetition, so the system harvests the creative tail of a model's distribution while keeping only hypotheses that survive grounding.
Gated Consensus Deferral
Subagents run a long-horizon research task and publish findings to a shared bulletin board, but gated isolation lets selected search branches keep their own working hypothesis private instead of reading peers' partial results immediately. Structured review only happens at a small number of explicit commitment boundaries, so only sufficiently confident candidates get propagated onward, and evidence gathering stays separated from evidence integration.
Repeated Labor Market Allocation
Agents bid on tasks using self-estimated private execution costs and self-maintained strategy notes; a central allocator selects winners from bids plus public reputation records and applies a VCG-style payment rule that rewards honest cost-aware bidding. Complex tasks are handled through hierarchical delegation: a winning agent can decompose its task and subcontract pieces back through the same bidding market rather than a planner assigning subtasks directly.
