Let models broaden the option set.

Video caption

Let models broaden the option set.

Nobody writes the observation that would end it.

Prediction, threshold, experiment, owner and dissent.

What I would require: Use the loop to produce tests, then test the system.

#EricFieldNotes

Full written post / accessible read

A pair of strong models can surface alternatives a busy team missed. That is a genuine advantage. The engineer's job is to supply the business consequence, prior incidents, acceptance threshold and test that can reject each attractive proposal.

Imagine two architectures for a risky workflow. Every model round produces another balanced comparison. If the team never names the failure cost or a discriminating test, the discussion consumes time while implementation waits. Consensus can become a substitute for a decision.

For each candidate, state what it predicts under the known failure case. Set the acceptable error or recovery time, run a safe test, and record observed evidence and unresolved dissent. The named owner chooses from that record, not from the debate's tone.

Ask agents to design rival probes and instrument the chosen one. Store the outcome and reason so future sessions do not restart the argument. Do this because model intelligence is most valuable when it accelerates evidence gathering, while engineers retain the authority to decide what evidence is enough.

#EricFieldNotes

Evidence and boundary

On-screen label: ILLUSTRATIVE DESIGN GATE. Illustrative cases are not measured incidents. Research papers and vendor documents support the stated mechanism only within their studied or documented scope.

• METR randomized trial (S20)
• Ki et al., Multiple LLM Agents Debate for Equitable Cultural Alignment (S78)
• METR 2026 productivity experiment update (S93)