An archive of decisions, failures, changes, and shipped work. The record stays useful by keeping the sequence—and the limits—visible.
Long-running agent recovery depends on preserving execution state, pending decisions, artifacts, and side-effect history rather than simply retrying the last request.
A retry is easy to reason about when the request is the whole unit of work. It fails, you try again, and — provided the operation is safe to repeat — nothing particularly complicated has happened.
A seven-layer diagnostic for tracing context failures upstream of the model window, from selection and representation to retrieval, compaction, permissions, inspection, and final capacity.
Why persistent AI memory needs a user-governance layer after the system has already decided what is valid enough to remember.
Why production evaluation needs its own maintenance contract when cases, judges, workload distributions, noise, and failure surfaces change.
Why some of the most important production agent failures need semantic evaluation and trace review because they never appear as system errors or explicit user feedback.
Why an agent that can modify its own harness needs a separate boundary between authoring the next candidate and trusting it as the next baseline.
What changes when a standardized agent interface can operate physical equipment: capability, authority, observation, recovery, and evidence become part of the execution contract.