Durable execution
Checkpoints, crash windows, replay, idempotency, retries, compensation, pause и resume.
Ограничения и автономность
Least agency, capability/policy boundaries, prompt injection, confused deputy и commit-time authorization.
Observability
Events, logs, metrics, traces, correlation IDs, redaction, latency, cost и replay.
Evaluation and data
Tasks, trials, outcomes, graders, repeated runs, judge calibration, sampling, leakage и data flywheel.
Production architecture
API, queue/workers, model gateway, state, tools, policy, retrieval, observability, evals и degradation.