Procurement should include hook failure injection.

Video caption

Procurement should include hook failure injection.

A permissive failure can become a production incident.

Safe denied action, injected faults and independent state.

My rule: A vendor event surface is only one part of the system.

#EricFieldNotes

Full written post / accessibility read

When evaluating an agent platform, I would ask for a live acceptance matrix: pre-call denial, post-call observation, local and cloud support, delegate behavior, timeout default and a target-state readback. 'Supports hooks' is not an architecture requirement.

If a hook process is missing on one runtime, will the agent proceed, stop, or ask a person? Can another tool reach the same service? A buyer needs those answers before granting production credentials, not after a pilot passes.

Give every candidate the same disposable repository and stateful canary sink. An authorized control must write a marker; denied routes must not. Inject timeout and malformed hook output, test each supported path, then verify the marker externally. Price the ongoing scripts and service permissions too.

Select the platform whose controls cover your real routes, and put the remaining authority in the target service. Do this because adoption speed is useful only when the team's rules survive tool errors, alternate paths and new product versions.

#EricFieldNotes

Four-beat scene transcript

1. Procurement should include hook failure injection.

When evaluating an agent platform, I would ask for a live acceptance matrix: pre-call denial, post-call observation, local and cloud support, delegate behavior, timeout default and a target-state readback. 'Supports hooks' is not an architecture requirement.

Visual: A feature checklist cannot tell you the control coverage.

2. The recovery cost belongs on the scorecard.

If a hook process is missing on one runtime, will the agent proceed, stop, or ask a person? Can another tool reach the same service? A buyer needs those answers before granting production credentials, not after a pilot passes.

Visual: A permissive failure can become a production incident.

3. Write acceptance as experiments.

Give every candidate the same disposable repository and stateful canary sink. An authorized control must write a marker; denied routes must not. Inject timeout and malformed hook output, test each supported path, then verify the marker externally. Price the ongoing scripts and service permissions too.

Visual: Safe denied action, injected faults and independent state.

4. Buy enforceable boundaries.

Select the platform whose controls cover your real routes, and put the remaining authority in the target service. Do this because adoption speed is useful only when the team's rules survive tool errors, alternate paths and new product versions.

Visual: A vendor event surface is only one part of the system.

Research and claim limits

• Cursor hooks documentation (S04)
• Codex hooks reference (S44)
• Claude Code hooks reference (S45)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.