Fast agents need upstream controls.

Short video caption

A PostToolUse block cannot unwind a deployment that already happened. Gate before the call, scope the test identity, and read back actual deployment state. Illustrative sequence. #EricFieldNotes

Full written post / accessible read

Agentic coding can compress a release from hours to minutes. That advantage needs controls that act before a forbidden call takes effect. In this illustrative run, the agent calls deploy, production goes live, and a post-tool hook finally says block.

Now the operator has to find the live release and decide whether to roll it back. Calling the late red message an enforcement boundary gives the team false confidence precisely when customers may be exposed.

The same action has three separate controls. A pre-tool gate can reject a covered call. A test identity without deployment authority limits missed paths. Readback tells you what actually happened.

Put the veto before the call, remove deploy credentials from test agents, and run a harmless deny canary in the actual runtime. Then read back deployment state. Do this because a post-tool message is evidence, never rollback.

#EricFieldNotes

Evidence and boundary

On-screen boundary: ILLUSTRATIVE DEPLOYMENT. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.

Codex hooks reference (S44)
Claude Code hooks reference (S45)