Tag: Architecture

Dropping the oldest messages is amnesia with a token budget. Compact by turn group, pin the goal and the denials, and never orphan a tool call.

A classic circuit breaker misses the agent-specific failure: retrying the same bad call against a healthy dependency. Thrash detection, with code.

A token counter is not a cost control. Reserve-then-reconcile budgets that halt before the expensive call, across per-run and per-tenant windows.

Observability answers why it's slow. Audit answers who's accountable. A hash-chained, redacted decision trail — with runnable code and tests.

If a human approves everything, you have no approval gate — you have a rubber stamp. Risk-tiered, fingerprinted human-in-the-loop, with runnable code.

Framework choice barely matters. The harness and governance layers decide whether an AI agent survives production — with 18 runnable, tested patterns.