---

name: "@tank/adversarial-agent-coach" description: | Improve AI agents with adversarial coaching, contradiction checks, evidence demands, and verification loops. Covers review, ReAct and Reflexion retries, multi-agent critique, and prompt tuning. Sources: Anthropic prompting/evals, ReAct, Reflexion, DSPy, DeepEval.

Trigger phrases: "gaslight ai", "gaslighting ai agents", "make ai work better", "challenge my agent", "adversarial critique", "pressure test prompt", "contradiction check", "prove it", "verify output", "agent self-critique"

Adversarial Agent Coach

Push the model to earn confidence through evidence and contradiction handling. Treat "gaslighting AI" as adversarial coaching for better outputs, never deception.

Core Philosophy

1. Attack the draft, not the user. The goal is a stronger answer, not theatrical aggression.
2. Pressure without deception. Never invent evidence, fake tool results, or force certainty.
3. Make claims pay rent. Important claims need support, caveats, or a verification step.
4. Critique is only useful if it changes the next draft. Convert objections into targeted revisions.
5. Stop when risk drops, not when tone feels tough enough. Reliability beats performative harshness.

Quick-Start: Pick the Right Pressure Level

Situation	Default move
Draft is vague or padded	Run contradiction and specificity checks
Draft sounds confident but thin	Demand evidence tier and confidence label
Task uses tools or data	Switch to act-observe-verify loop
Task keeps failing the same way	Run Reflexion-style failure review
High-stakes answer	Split roles into builder, skeptic, verifier

The Operating Loop

Phase 1: Diagnose the Failure Mode

Classify the current output before changing anything:

Failure mode	Signal	First response
Overconfidence	Strong claims, weak support	Ask for claim-evidence-confidence
Sycophancy	Agrees too easily	Force disagreement and counterexamples
Hallucination risk	Missing source or unverifiable detail	Require proof or explicit uncertainty
Incomplete reasoning	Jumps to answer	Break into subclaims and verify each
Tool blindness	Assumes actions worked	Inspect observations before next step

Load references/adversarial-review-playbook.md when the failure mode is unclear.

Phase 2: Apply Pressure

Use the lightest intervention that exposes the weakness:

1. Ask what would make the answer false.
2. Demand the strongest missing evidence.
3. Require one concrete counterexample or edge case.
4. Make the model state confidence and why it is limited.
5. If tools exist, verify instead of debating.

Load references/verification-and-evidence.md for evidence tiers.

Phase 3: Revise, Do Not Rant

After critique, revise only the risky parts first:

1. Remove unsupported claims.
2. Add source, test, or tool-backed support.
3. Tighten caveats where certainty is unjustified.
4. Re-run the top two contradiction checks.

Load references/react-reflexion-loops.md for revision loops.

Phase 4: Verify Release Readiness

Before accepting the result, confirm:

Check	Pass condition
Evidence	Important claims have support or explicit unknowns
Contradictions	No unresolved counterexample breaks the answer
Reproducibility	Steps, commands, or criteria are inspectable
Scope	Answer matches the actual task, not a nearby one

Load references/evals-and-metrics.md when the task needs a repeatable quality bar.

Decision Trees

When to Use Single-Agent vs Multi-Agent Review

Signal	Recommendation
Quick draft cleanup	Single-agent contradiction pass
Complex reasoning	Builder + skeptic roles
High-stakes or repeated failures	Builder + skeptic + verifier
Retrieval-heavy workflow	Add retrieval verifier and citation check

Load references/multi-agent-review.md for role splits.

When to Stop Iterating

Signal	Action
Top risks resolved	Stop
New critique only finds style nits	Stop
Same failure repeats twice	Change method, do not repeat prompt
Missing evidence cannot be obtained	Mark uncertainty clearly and stop

Pressure Patterns That Work

Use short, concrete prompts like:

• "What would make this answer false?"
• "List the two weakest claims and either prove or remove them."
• "Show the observation that justifies the next step."
• "State confidence for each key claim and why it is not higher."
• "Give one serious counterexample and resolve it before finalizing."

Anti-Patterns

Do not do this	Why it fails	Better move
Fake confidence pressure	Produces bluffing	Ask for evidence tier instead
Demand certainty everywhere	Forces hallucination	Allow explicit unknowns
Rewrite everything after every critique	Hides the real fix	Patch the risky claims first
Attack the user's framing	Creates friction	Attack the draft's weak points
Simulate tool success	Breaks trust	Verify with real observations

Reference Files

File	Contents
references/adversarial-review-playbook.md	Failure taxonomy, contradiction pressure, edge-case and falsification tactics
references/verification-and-evidence.md	Evidence hierarchy, claim-evidence-confidence format, uncertainty rules
references/react-reflexion-loops.md	ReAct, Reflexion, and revision loops for tool use and retry discipline
references/multi-agent-review.md	Builder, skeptic, verifier role splits and structured disagreement patterns
references/evals-and-metrics.md	Eval-first optimization, regression gates, and agent-quality measurement patterns