Event and knowledge bus

Communication should be explicit

A model ecology needs communication, but not every model needs direct access to every other model. A typed event and knowledge bus creates stable contracts, replayable traces, and permission boundaries.

Event classes

Knowledge objects

Knowledge objects are versioned artifacts such as retrieval corpora, embeddings, distilled datasets, taxonomies, calibration maps, or failure clusters. They have manifests, provenance, access rules, and expiration. A shared vector store without these properties can become an untracked cross-model memory channel.

Event envelope

event <- {
    event_id: UUID_V7(),
    event_type: "evaluation.completed",
    occurred_at_utc: NOW_UTC(),
    trace_id: trace_id,
    producer_id: evaluator_service_id,
    subject_id: candidate_artifact_id,
    schema_version: "1.2",
    data_classification: "internal",
    payload_hash: HASH(payload),
    payload: payload
}

Delivery semantics

Choose at-least-once or exactly-once behavior consciously. Make consumers idempotent using event identifiers. Critical state transitions should not depend solely on a transient message; persist a decision record before publishing the event.

Avoiding conversational cascades

Free-form multi-agent debate can amplify errors, create hidden state, and make causality difficult to reconstruct. Prefer independent generation, typed intermediate artifacts, bounded rounds, and explicit aggregation. When conversational coordination is necessary, log every message and enforce a strict turn and token budget.

Security

Authenticate producers, authorize event types, validate schemas, limit payload size, and prevent untrusted output from changing routing or policy. Treat model-generated strings as data, not commands to downstream services.