Specialist agents may reason, exchange findings and coordinate in parallel. That does not mean every agent should acquire permission to approve, publish, spend or change consequential state.

Architecture principle: Coordination topology describes how work and information move. Authority topology describes who may authorize consequential state transitions. These are separate design decisions.

Two graphs, not one hierarchy

Multi-agent systems are often framed as a choice between one central supervisor and autonomous peers. A single scheduler can become a bottleneck; a peer-to-peer team can confuse self-assignment with authorization. Neither problem is solved by assuming that coordination and authority must have the same structure.

The L3 runtime can decentralize task discovery, bounded execution, artifact exchange and reconciliation. Separately, the authority graph governs policy decisions, qualification, approval, release and externally visible effects. Central authority means consistent, accountable decision rights, not one central message broker or a human approval prompt for every harmless step.

COORDINATION: worker A <-> shared workspace <-> worker B
                         | proposals / evidence |
                         v
AUTHORITY:    policy -> qualification -> approval -> effect gate
                                                  |
                                                  v
                                         observed effect receipt

The primary layer is L3 Runtime & Orchestration. It has two explicit interfaces: L6 Security, Identity & Governance supplies enforceable authorization, while L7 Evaluation & Observability supplies independent qualification and effect evidence. Delegation, progression and synthesis remain L3 responsibilities.

What the research demonstrates and what it does not

Agensh (Zhan et al., September 2026) describes self-organizing workers sharing a workspace, messages and reusable context. In the authors' five selected ProgramBench tasks, mean final test-pass rate increased from 19.31% with one agent to 28.78% with 128. A separate Pandoc experiment reported 33.89% with one worker and 55.06% with 1,024. These are author-reported, task-specific preprint findings. They do not isolate decentralization as the cause of improvement, hold total compute constant, or establish enterprise deployment safety.

Raven (EverMind AI, September 2026) explores a host that composes heterogeneous model-harness execution units, validates dependency graphs, routes artifacts and synthesizes results. Its useful architectural contribution is to treat a specialist as a complete runtime contract, not merely a model name or a prompt role. Its evaluations do not independently validate the control architecture proposed here.

The engineering inference is deliberately narrower than the papers' performance claims: distributed reasoning may be useful, but task assignment, peer agreement and local benchmark success are not policy approval.

Contract the specialist, not merely its role

A specialist is a versioned execution unit: model revision, tools, context policy, memory behavior, execution limits and output obligations. A delegation envelope should preserve:

  • Identity and lineage: parent task, worker identity, accountable sponsor and execution-unit revision.
  • Bounded objective: requested deliverable, permitted inputs, deadline, resource budget and stopping condition.
  • Authority: allowed read and draft operations, prohibited effects and requirements for separate approval.
  • Artifact contract: input and output schema, source references, provenance, acceptance criteria and error states.
  • Lifecycle: lease and expiry, cancellation, retries, idempotency, checkpoints and recovery ownership.

Tool discoverability is not permission. A worker's completion signal is not proof that the requested outcome occurred. Mutable shared-workspace records must not silently replace historical evidence.

From parallel recommendations to one effect decision

Consider three specialists assessing a software release: one checks tests, one reviews security findings and one drafts release notes. They can work concurrently and return contradictory findings. The synthesizer should preserve the disagreement and propose a decision packet. It must not convert a majority vote or an apparently passing test into permission to deploy.

Before a consequential operation, the final gate resolves the actual artifact, target, principal, task, scope and applicable policy. It checks whether an accountable approval binds that exact operation. A material change after review requires a new decision. After execution, an independent observer checks whether the intended target state actually changed.

Six adverse-condition tests

  1. Duplicate claim: two workers claim the same task; leases and state versions prevent duplicate effects.
  2. Stale approval: the target or arguments change after review; the effect is rejected or reapproved.
  3. Conflicting evidence: specialists disagree; synthesis retains both claims and source provenance.
  4. Unavailable evaluator: qualification remains incomplete and progression stops.
  5. Worker timeout: retry preserves task lineage, budget and delegated authority limits.
  6. Partial deployment: observed target state triggers recovery rather than a success claim.

These are proposed conformance tests, not claims that the controls are deployed or independently validated. Compare the design against a simpler topology on the same task population and disclosed resource budget. Measure qualified outcomes, end-to-end cost, time to completion, handoff errors, denied effects and recovery quality, not agent count alone.

Decision, evidence and limitations

Architecture decision: permit decentralized low-consequence coordination within explicit specialist contracts while keeping consequential progression and effects subject to independently enforceable policy, qualification and accountable approval. Prefer the least complex topology that meets measured requirements.

Evidence boundary: Agensh supports feasibility of asynchronous collaboration in selected software tasks. Raven motivates typed composition of heterogeneous execution units. NIST AI RMF 1.0 supports general accountability and measurement but does not prescribe this two-graph architecture. None of these sources proves universal scaling gains or production safety.

Open questions: When does coordination overhead outweigh benefit? How can independent evaluation avoid contamination from shared context? Which bounded actions can be delegated without repeated human approval? What evidence proves that a final effect matched the reviewed operation through retries?

Related reading: Promotion Is a Control (L2) · The Harness Is Part of the Model (L3 foundation).

Privacy Notice: With your consent, this website uses Google Analytics to understand aggregate visits and content usage. Analytics is optional and is not used for advertising or cross-site profiling. Some content and images may be generated or modified with the assistance of AI; every AI-assisted item is reviewed and manually edited by a human before publication. Full Privacy Policy · AI modified — reviewed and edited by a human