Where state failures hurt — and how BinSentry fixes them.

An agent calls a flight-search tool, caches the result for 'speed', and reuses it 40 minutes later when prices have already moved. The user books a fare that no longer exists.

• TTL policy on the `flight.search` bin caps freshness at 90 seconds.
• A freshness gate blocks reads past TTL and forces a re-fetch.
• Every blocked read is emitted as a telemetry event so you can tune TTLs by tool.

A multi-step research agent produces a wrong answer in production. By the time an engineer looks at the trace, the scratch files, browser snapshots, and tool outputs are gone.

• BinSentry persists every bin with a content hash for the run's retention window.
• Replay a run from any step with the exact original artifacts.
• Diff two runs to see which bin changed and why the outcome flipped.

Your agent touches PII while drafting a customer reply. Legal asks: which data, from which source, ended up in which output, on which date?

• Sensitivity tags propagate through every derived bin automatically.
• Lineage graph answers 'what fed this output?' as a single query.
• Residency + redaction policies are enforced before bins are written.