Interview for the work agents leave humans.

Video caption

Interview for the work agents leave humans.

The boundary should be substantive.

Same artifact, questions, facts and score anchors.

My rule: The rubric itself needs evidence.

#EricFieldNotes

Full written post / accessibility read

Before designing an AI-era interview, map the decisions this role really makes. Which architecture, verification, release and handoff judgments recur? Then ask a few open, role-related questions over the same short artifact. OPM supports job analysis and structured scoring; this particular three-question format is my proposal, not a validated universal test.

Show a fictional event migration: producer tests pass, consumer compatibility remains unresolved. Ask what the candidate would need to decide before release. A strong answer names the external owner, replay risk, rollout sequence and stop condition. The interviewer should prepare those facts and scoring anchors; the candidate should not have to construct the system.

Pilot the short interview with experienced engineers. Give each candidate the same artifact and questions, then reveal the same late consumer constraint. Have two trained reviewers score a sample independently and compare disagreements. Use a code-reading follow-up for data flow, concurrency, security and operations; rhetoric should not stand in for competence.

Track whether these questions predict accepted changes, review quality and useful handoffs after hiring, where appropriate and lawful. Do this because the scarcity moves toward decisions around code, but a polished interview theory should not become another untested process. Keep the core engineering check and revise the rubric from observed outcomes.

#EricFieldNotes

Four-beat scene transcript

1. Interview for the work agents leave humans.

Before designing an AI-era interview, map the decisions this role really makes. Which architecture, verification, release and handoff judgments recur? Then ask a few open, role-related questions over the same short artifact. OPM supports job analysis and structured scoring; this particular three-question format is my proposal, not a validated universal test.

Visual: Start from the actual role, not an AI slogan.

2. Make the diff look plausible.

Show a fictional event migration: producer tests pass, consumer compatibility remains unresolved. Ask what the candidate would need to decide before release. A strong answer names the external owner, replay risk, rollout sequence and stop condition. The interviewer should prepare those facts and scoring anchors; the candidate should not have to construct the system.

Visual: The boundary should be substantive.

3. Calibrate the questions before candidates arrive.

Pilot the short interview with experienced engineers. Give each candidate the same artifact and questions, then reveal the same late consumer constraint. Have two trained reviewers score a sample independently and compare disagreements. Use a code-reading follow-up for data flow, concurrency, security and operations; rhetoric should not stand in for competence.

Visual: Same artifact, questions, facts and score anchors.

4. Hire for accepted work, not diff theater.

Track whether these questions predict accepted changes, review quality and useful handoffs after hiring, where appropriate and lawful. Do this because the scarcity moves toward decisions around code, but a polished interview theory should not become another untested process. Keep the core engineering check and revise the rubric from observed outcomes.

Visual: The rubric itself needs evidence.

Research and claim limits

• U.S. OPM: Work Samples and Simulations (S151)
• U.S. OPM: Designing an Assessment Strategy (S156)
• Vella and Blincoe, longitudinal AI coding-assistant study (S154)
• U.S. OPM structured interviews (S203)
• U.S. OPM structured interview guide (S204)

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