Agentic AI and multi-agent stack monitoring

Supervisor pattern: separate planning from policy. A coordinator may delegate to specialised workers; a deterministic policy layer must authorise data access and effects.

Communication and state: use typed messages, an immutable event history and explicit workflow state; conversational memory is not a transaction record.

Memory and RAG: separate user-approved durable memory from the task scratchpad, and attach source, time, consent and retention metadata.

Human approval: bind approval to the exact action, recipient, payload and expiry; replanning must invalidate old approvals.

Recovery: keep idempotent effects, compensation, bounded retries and dead-letter review; provenance at checkpoints must carry no secrets.

Audit: log the model and version, route, tool arguments, policy result, approver, cost, latency and external effect.