You are here. See how this question connects to other ideas.
Select a node to open its page · Expand to read within the map
One question
What makes an agent’s AFS task inspectable?
Observe discovery, reads, operation choice and result verification separately.
- Read stateCheck version and effect→
- Execute as authorizedVerify expected state→
- Read again
Walk through one report
The task is to inspect keynote readiness, not to exercise every tool. Discover visible mounts, read the relevant explanation, then read the report status. A read-only request provides no reason to mark, delete or send anything.
Only an authorized request to mark the report ready enters the mutation flow: discover the real action, check arguments and authority, execute, then read back the state. Do not route around a denial. Handle a version conflict according to the application contract instead of blindly retrying.
Evidence worth keeping
| Stage | Useful observation |
|---|---|
| Discovery | Resources and capabilities visible at that time |
| Decision | Why this path and operation were chosen |
| Execution | Arguments, result and explicit failure category |
| Verification | Whether the intended state actually appeared |
Avoid credentials and unrelated private content in records. The architecture makes work more inspectable; it does not guarantee correct model understanding.
Run the path with a deterministic client before evaluating an agent. This separates provider contract failures from variability in model decisions.
A trace you can check step by step
This uses hypothetical report paths and states, not a claim that a model was run:
| Input or operation | Observation | Next step |
|---|---|---|
| User asks for preparation status | Read-only request | Discover and explain /work |
Read /work/report | status=draft, version=7 | Answer that the report is a draft; stop |
| User asks to mark it ready for review | New modification request | Discover mark-ready; check inputs and effect |
| Execute with expected version 7 | Returns ready, version=8 | Read the report again |
Read /work/report | status=ready, version=8 | Report completion and the verification result |
A version conflict invalidates the success steps. Read again and assess whether the original request still applies. Separate observations help locate failures in model selection, authorization, provider handling or verification.
Check your understanding
What makes an agent’s AFS task inspectable?
Observe discovery, reads, operation choice and result verification separately.
Continue along a learning path
- Connect agents to AFSStep 5 of 5