FIELD NOTE / INSTAGRAM
Turn standards into checkpoints.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved Instagram edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
Turn standards into checkpoints.
Video caption
Turn standards into checkpoints.
Post-call feedback can arrive after the write.
Scripted gate, least privilege and protected acceptance.
Record action, rule, denial and post-state.
#EricFieldNotes
Full written post / accessibility read
A team's architecture guide can say 'use the tenant-scoped repository.' If an agent can still call a raw database tool with broad credentials, the guide is advice. The checkpoint belongs at the tool boundary and the final accepted-state test.
In a simulated migration, an agent performs a broad update and the post-tool hook reports a violation afterward. The message is useful audit evidence. It does not undo the write. A real control needs the right ordering and a credential that limits missed paths.
A Python guard parses the proposed tool input before use. The agent's database identity cannot write outside its tenant. A separate CI job runs tenant-isolation fixtures after the patch. Inject a deliberate cross-tenant attempt and require all three to show the expected behavior.
Attach the canary result and protected test report to the change review. Do this because an agent can write plausible compliance prose, while an executable checkpoint leaves a verifiable reason the bad path could not complete.
#EricFieldNotes
Four-beat scene transcript
1. Turn standards into checkpoints.
A team's architecture guide can say 'use the tenant-scoped repository.' If an agent can still call a raw database tool with broad credentials, the guide is advice. The checkpoint belongs at the tool boundary and the final accepted-state test.
Visual: The agent should meet the rule at the moment it can matter.
2. A late warning is not prevention.
In a simulated migration, an agent performs a broad update and the post-tool hook reports a violation afterward. The message is useful audit evidence. It does not undo the write. A real control needs the right ordering and a credential that limits missed paths.
Visual: Post-call feedback can arrive after the write.
3. Use three enforceable layers.
A Python guard parses the proposed tool input before use. The agent's database identity cannot write outside its tenant. A separate CI job runs tenant-isolation fixtures after the patch. Inject a deliberate cross-tenant attempt and require all three to show the expected behavior.
Visual: Scripted gate, least privilege and protected acceptance.
4. Make violation evidence reviewable.
Attach the canary result and protected test report to the change review. Do this because an agent can write plausible compliance prose, while an executable checkpoint leaves a verifiable reason the bad path could not complete.
Visual: Record action, rule, denial and post-state.
Research and claim limits
- Cursor hooks documentation (S04)
- OpenAI plugin architecture (S06)
- OpenAI Agents SDK/Codex orchestration example (S07)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.