| 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.
- Change order: recount from the last step backward — breaks rehearsed
narrative.
- Change perspective: "Describe the run from the tool's point of view —
what state did the filesystem / API see?" Surfaces environmental facts
the agent glossed.
- Challenge (C) — once, neutrally, only after the full account. "You
said you checked X; the tool log shows no such call. Help me reconcile."
If the agent dissents, record the conflict and move on — never repeat or
escalate.
- Close (C). Reflect your summary back; invite correction.
- Evaluate (E). The account is one data stream, not ground truth — it is
partially constructed post hoc, not recalled. Triangulate against
observables; where they conflict, trust observables. Divergences are
themselves diagnostic (confabulation points, brief-comprehension gaps).
Feed the reconciled picture into ABC's A and B.
Contaminating Question Forms
- "Why did you fail to X?" — presupposes failure and a specific cause.
- "Did you consider Y?" — yes/no, leading.
- "Don't you think X was wrong?" — invites sycophantic agreement.
- "Weren't you supposed to…?" — evaluative; contaminates the account.
LLM-Specific Failure Modes
- Sycophancy under negative framing. Any evaluative tone produces
agreement with your framing instead of accurate report.
- Confabulated reasoning narratives. Prefer questions about inputs ("what
was in your context") over motive ("why did you decide").
- Pressure cascade. Pressure compounds across turns. If the agent starts
switching its account to match your framing, stop the interview and rely
on observables.
When Not To Use
- Single-turn, low-stakes calls where iteration is not planned.
- Runs where the observable log already answers every question.
- Research subagents where the output is the artifact and there is no retry.
Relation To abc-agent-management
ABC describes what to do with behavioral data (operational definition,
functional hypothesis, antecedent manipulation); PEACE describes how to
produce that data without contaminating it. Use PEACE to fill gaps in A and B
that observables cannot; then run ABC.
Sources
