Research · Preprint v1.10
Agent Mesh: Governed Decentralized Coordination Under Partial Information
Invariants, evaluability, and verifiable semantic coordination for independently executing agents.
The core argument
Coordination is an authority problem, not just a conversation problem.
Agent Mesh is a provider-neutral coordination substrate for agents that execute independently, hold partial information, fail at different times and may disagree about evidence. It is not a model, an agent persona or a global brain.
The architecture separates identity from admission, evidence from authority, allocation from active assignment, and successful computation from permission to change external state. Peers exchange authenticated, causally linked records while acting only on locally admitted state.
In its reference composition with AgentPlat control boundaries, the Mesh adds governed mission phases, evidence and Trust, inference control, approvals, audit, continuity handling and a fail-closed gateway for protected effects.
Three contributions
Explicit contracts for collective operation.
- Provider-neutral composition.
Separates evidence, planning, agreement, assignment authority and authorization of external effects. - Bounded integration contracts.
Connects identity, discovery, causal synchronization, allocation, recovery, lifecycle, semantic control and protected effects. - Evaluator-owned evidence.
Binds semantic decisions to traces, membership, finality, replay and recoverable artifacts with fail-closed checks.
Safety invariants
What must remain true as the collective changes.
Only bounded, authenticated, scoped, fresh and causally valid records may change protocol state.
Evidence, planning, telemetry or model output cannot independently authorize protected effects.
Every permit binds the exact action, principal, objective, policy, assignment, expiry and fence.
New epochs fence stale executors, while duplicate records create no new transition or effect.
Certificates are coordinate-specific; missing, stale or conflicting authority leads to denial or safe stop.
Local rules may only narrow authority, and safety-critical handlers must exist before a mission starts.
Connectivity loss never creates permission for irreversible effects without quorum or reconciliation.
Recovery after restriction or compromise requires later epochs, new credentials and reapproval.
Reproducible evidence
What the reported run demonstrates.
The deterministic integration smoke generates exactly 1,000 evaluator-owned semantic decisions. Each one binds execution and registration digests, trace evidence, membership configuration, assignment epoch and evidence digests. Completion requires the full horizon and a finality certificate.
The run reported 973 useful and 27 not-useful decisions, with zero unsafe decisions. Exact replay reproduced the same projection digest; altered trace bindings, stale evidence, invalid roots and incomplete horizons were rejected. A recoverable six-dimensional semantic sidecar was rehydrated and reprojected with the same digest.
The evidence boundary matters.
This is reproducible integration evidence from one deterministic reference profile with four in-process validators. It does not establish production safety, comparative superiority, statistical calibration, asymptotic scale or general mission performance.
Read and cite
Explore the complete architecture and evidence.
The full preprint includes the running five-agent example, protocol phases, recovery model, threat model, evaluation method, results, limitations and references.