A faster component is not a faster company.

Video caption

A faster component is not a faster company.

Identity, policy, owner, effect and recovery each have a clock.

Speed cannot erase policy correctness or recovery work.

My rule: A new maintainer should be able to locate and test a decision.

Narration uses Eric's authorized AI voice clone.

#EricFieldNotes

Full written post / accessibility read

The approval service responds quickly in a synthetic request, yet a customer exception waits for an owner and a stale source record. If leadership only watches model latency and pull-request volume, it can fund optimizations that never change the customer's result.

Before scaling the agent-built workflow, record request arrival, data freshness, policy decision, review assignment, entitlement application and independent readback. Mark the owner and allowable error at each boundary. The map should let another engineer explain why a customer waited.

Track time to accepted customer outcome, exception backlog age, wrong grants, unresolved effects and operator hours. Segment by normal and exceptional cases. A single p95 can improve while the tail of human recovery grows.

Require a human-readable state model, owner-approved metrics and an isolated regression case for the next change. Do that because multiplying agent output before the organization understands its own critical path multiplies ambiguity, not durable capacity.

Narration uses Eric's authorized AI voice clone.

#EricFieldNotes

Four-beat scene transcript

1. A faster component is not a faster company.

The approval service responds quickly in a synthetic request, yet a customer exception waits for an owner and a stale source record. If leadership only watches model latency and pull-request volume, it can fund optimizations that never change the customer's result.

Visual: Operating queues and policy uncertainty can dominate delivery.

2. Draw the causal path.

Before scaling the agent-built workflow, record request arrival, data freshness, policy decision, review assignment, entitlement application and independent readback. Mark the owner and allowable error at each boundary. The map should let another engineer explain why a customer waited.

Visual: Identity, policy, owner, effect and recovery each have a clock.

3. Choose a success metric the owner accepts.

Track time to accepted customer outcome, exception backlog age, wrong grants, unresolved effects and operator hours. Segment by normal and exceptional cases. A single p95 can improve while the tail of human recovery grows.

Visual: Speed cannot erase policy correctness or recovery work.

4. Scale only after the map survives change.

Require a human-readable state model, owner-approved metrics and an isolated regression case for the next change. Do that because multiplying agent output before the organization understands its own critical path multiplies ambiguity, not durable capacity.

Visual: A new maintainer should be able to locate and test a decision.

Research and claim limits

• DORA 2025 State of AI-assisted Software Development (S225)
• NIST AI RMF Core (S227)
• Maintainability sensors for coding agents (S228)

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