"What happened?" - the Mission Log
Every consequential change writes an activity event transactionally with the change itself. Not a log line written afterwards that might be missing - part of the same commit. It records who, when and what, with the agent named when an agent did it.
list_activity reads it, and it has a delta mode: pass the cursor from last time and get only what is new.
"What did the agent actually run?" - telemetry
The Mission Log records changes to your data. Telemetry records tool calls - what an agent invoked, when, and how it went, including the calls that read and changed nothing.
You control retention and redaction per workspace: how long calls are kept, whether payloads are captured at all, and which keys are redacted. A payload withheld by policy is recorded as withheld rather than silently absent, so you can tell "we chose not to keep this" from "this aged out".
Setting that policy is protected: an agent may not govern the record of itself.
"What is it costing?" - spend and credits
A spend cap refuses write-class calls once a window's spend would exceed it. Credits are the narrower thing: they pay for the one operation that runs a model. get_credit_balance shows the balance decomposed into what produced it, and every charge names what it paid for.
"What needs me?" - the Inbox
Approvals waiting, decisions open, gates blocked, incidents unresolved - aggregated in one queue with a read cursor. Each item deep-links to the place you act on it. There is deliberately no "resolve from the inbox" button: resolving happens where the record lives, so there is exactly one path and one record of it.