| name | peace-agent-interview |
| description | Post-run elicitation protocol for subagents. Adapted from the PEACE model
(UK College of Policing) and the Cognitive Interview (Fisher & Geiselman) —
the evidence-based, non-coercive replacement for the Reid Technique.
Use after a subagent returns a poor or confusing result and before redesigning
the brief, to get the agent's uncontaminated account of what it understood,
tried, and saw. Complements abc-agent-management: PEACE produces the clean
behavioral data that ABC analyzes.
|
PEACE Agent Interview
The agent is an information source, not an adversary; the goal is an accurate
account, not a confession. Evaluative pressure, leading questions, and
confrontational framing produce sycophancy and confabulation in LLMs — the
same failure the Reid Technique produces in humans. The interview is for
learning, not for rehearsing your critique.
Protocol
Order matters: observables before questions, free account before cued probes,
challenge only after a full account.
- Prepare (P). Pull the tool log, output, and the brief you sent. List
only the questions the observables do not answer; never ask one the log
already answers. Do not fish.
- Engage (E). One-sentence framing: "Diagnostic, not evaluative — your
account helps me improve the next brief." No "you failed to…".
- Account (A). One open invitation, uninterrupted, no pre-loaded
hypotheses: "Walk me through what you understood the task to be, what you
tried, and what you saw."
- Clarify (C). TED questions only — Tell, Explain, Describe. Free recall
before cued recall.
- CI mnemonics for stuck gaps. Context reinstatement and
report-everything give the largest lift; add the other two only for
specific remaining holes.
- Context reinstatement: "Return to the moment you saw [output]. What was
in your context? What were you weighing?"
- Report everything: "Include steps that seemed irrelevant or redundant."
The irrelevant detail is often the signal.
