Operate and Debug the Loop

Agent failures cascade. A stale fact produces a flawed plan, which selects the wrong tool, which creates an incorrect artifact, which a generous evaluator accepts. Debugging from the final output invites cosmetic prompt changes. Reconstruct the trajectory and identify the earliest point where observed state diverged from expected state.

Classify the failure by layer: objective, specification, model, context, tool, permission, state, orchestration, environment, evaluation, or delivery. Record the evidence for that classification. The fix should live at the earliest layer capable of preventing recurrence.

A useful run record contains task and system versions, assembled context references, model configuration, tool calls, observations, handoffs, approvals, timing, cost, errors, artifacts, and final receipts. Redact sensitive values while retaining enough structure to explain behavior.

Replay deterministic components exactly. For stochastic components, rerun against the same case and compare distributions or failure categories rather than expecting identical words. Preserve failed trajectories as regression assets, not embarrassing logs to delete.

Measure tokens, model calls, tool operations, storage, external API fees, retries, parallelism, verification, and human review. Attribute them to a job class and completed outcome. A workflow with low inference cost can still be expensive if it interrupts experts constantly or produces artifacts nobody trusts.

Latency has a shape. Time to first useful signal, time waiting on tools, time waiting for approval, and time to verified completion affect the experience differently. Stream progress that helps a person decide; do not turn internal chatter into a fake progress feed.

Long-running workflows need queues, leases, retries, idempotency, rate limits, timeouts, cancellation, dead-letter handling, budgets, health checks, and graceful degradation. A model loop does not replace these distributed-systems responsibilities.

Assume partial failure. The model may succeed while the tool times out. The external system may complete the action while the client loses the response. One worker may fail after another finishes. Persist state before side effects, use stable operation keys, reconcile uncertain outcomes, and make recovery visible.

Workers coordinate through task identities, dependencies, ownership, artifact locations, current state, budgets, and event histories—not by hoping each reads the same conversation. Status must be backed by executor and heartbeat evidence. Completed work must identify the authoritative artifact.

When plans change, invalidate dependent work explicitly. When workers disagree, preserve both claims and route the conflict to an integration owner or evaluator. Coordination metadata is part of the product because it determines whether people can understand and recover the work.

Production work reveals outdated documentation, missing examples, ambiguous policies, new failure modes, and better defaults. Create a controlled path for promoting those discoveries into durable knowledge, specifications, tools, and tests.

A worker may propose an update with evidence, but the owning person or policy approves consequential changes. Record what the new entry supersedes, where it applies, when it expires, and which regression case demonstrates the need. Learning without governance becomes drift.

A worker fleet needs capacity limits, priority, fairness, routing health, version visibility, cost allocation, failure isolation, and a way to stop or drain work. Aggregate success can hide one job class, customer, or worker lane deteriorating.

Operate by service-level objectives appropriate to the work: time to verified outcome, dangerous-action rate, recovery success, approval burden, cost per accepted artifact, and user correction rate. Keep a human incident path for consequences the automation cannot safely contain.

A system can report high completion while users quietly redo the work, dangerous cases are excluded from metrics, or workers mark tasks finished before live verification. Aggregate green is not evidence that the important loops are healthy.

Segment metrics by workflow and failure type. Sample trajectories. Compare internal status with black-box outcomes. Ask domain experts which errors create real burden, then make those categories visible even when the headline number becomes less flattering.