As code volume rises, judgment gets scarcer.

Video caption

As code volume rises, judgment gets scarcer.

Attention drift turns generated volume into hidden risk.

Use a messy case with incomplete facts and a test budget.

My rule: Keep fundamentals, add focus and clear decisions.

#EricFieldNotes

Full written post / accessibility read

Agentic SDLC can make implementation cheaper while multiplying choices about architecture, customer promises, exceptions and QA. The engineer who can focus, frame a problem, challenge a fluent answer and communicate a decision becomes more valuable. Fundamentals still matter: without them, the challenge is shallow.

An agent proposes a confident migration plan. A busy reviewer uses another model to summarize it. The summary drops the only legacy-state exception. Everyone feels informed, yet nobody has inspected the evidence or named the product owner. A lower hallucination rate would not fix the missing handoff.

Give a candidate a working agent-generated feature plus an ambiguous customer rule. Ask them to identify the unknowns, choose two discriminating experiments, allocate agent tasks, protect a release gate, and explain the tradeoff to product. Score their causal model and follow-through, not raw keystrokes.

Reward candidates who can understand the system, stay attentive across parallel work, refuse a convenient untested claim, and own an observed outcome. Do this because agents can supply more implementation than a team can safely absorb without engineering judgment.

#EricFieldNotes

Four-beat scene transcript

1. As code volume rises, judgment gets scarcer.

Agentic SDLC can make implementation cheaper while multiplying choices about architecture, customer promises, exceptions and QA. The engineer who can focus, frame a problem, challenge a fluent answer and communicate a decision becomes more valuable. Fundamentals still matter: without them, the challenge is shallow.

Visual: Hiring for typing speed misses the new bottleneck.

2. A plausible answer tempts passive review.

An agent proposes a confident migration plan. A busy reviewer uses another model to summarize it. The summary drops the only legacy-state exception. Everyone feels informed, yet nobody has inspected the evidence or named the product owner. A lower hallucination rate would not fix the missing handoff.

Visual: Attention drift turns generated volume into hidden risk.

3. Interview for decision ownership.

Give a candidate a working agent-generated feature plus an ambiguous customer rule. Ask them to identify the unknowns, choose two discriminating experiments, allocate agent tasks, protect a release gate, and explain the tradeoff to product. Score their causal model and follow-through, not raw keystrokes.

Visual: Use a messy case with incomplete facts and a test budget.

4. Hire people who can steer and verify.

Reward candidates who can understand the system, stay attentive across parallel work, refuse a convenient untested claim, and own an observed outcome. Do this because agents can supply more implementation than a team can safely absorb without engineering judgment.

Visual: Keep fundamentals, add focus and clear decisions.

Research and claim limits

• METR randomized trial (S20)
• Stack Overflow 2025 developer survey (S21)
• METR 2026 productivity experiment update (S93)

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