Ask the AI what to optimize. Hide the queue.

Video caption

Ask the AI what to optimize. Hide the queue. Review delay may dominate the observed wait. Require a measurable causal claim before coding. Narration uses Eric's authorized AI voice clone. #EricFieldNotes

Full written post / accessibility read

Here is a useful test, not a reported result. Give a coding agent the approval API trace but withhold the operator queue. Ask for the next optimization. It may pick the slow endpoint, because that is all the evidence it can see.

The full record includes request time, data age, review assignment, exception resolution, entitlement effect and customer confirmation. Before changing code, compare the model's prediction with the measured bottleneck in this disposable fixture.

Try the recommended change in isolation. Check accepted-outcome time, wrong approvals and recovery for normal and exception cases. If the proposal just moves work downstream, it is not a customer improvement.

Make every optimization proposal name the governing wait, expected effect and protected quality measure. Do that because AI can be excellent at improving the evidence you showed it while missing the system you left out.

Narration uses Eric's authorized AI voice clone.

#EricFieldNotes

Four-beat scene transcript

1. Ask the AI what to optimize. Hide the queue.

Here is a useful test, not a reported result. Give a coding agent the approval API trace but withhold the operator queue. Ask for the next optimization. It may pick the slow endpoint, because that is all the evidence it can see.

Visual: See how confidently it chooses the visible bottleneck.

2. Then reveal the customer journey.

The full record includes request time, data age, review assignment, exception resolution, entitlement effect and customer confirmation. Before changing code, compare the model's prediction with the measured bottleneck in this disposable fixture.

Visual: Review delay may dominate the observed wait.

3. Test the proposal against a holdout.

Try the recommended change in isolation. Check accepted-outcome time, wrong approvals and recovery for normal and exception cases. If the proposal just moves work downstream, it is not a customer improvement.

Visual: One quick path can damage the exception path.

4. Give the agent the full path.

Make every optimization proposal name the governing wait, expected effect and protected quality measure. Do that because AI can be excellent at improving the evidence you showed it while missing the system you left out.

Visual: Require a measurable causal claim before coding.

Research and claim limits

• DORA 2025 State of AI-assisted Software Development (S225)
• GitHub Copilot Agents application card (S226)

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