devils-advocate

Devil's Advocate — Adversarial Review via Subagent

Overview

Self-critique fails because the same biases that produced the proposal produce the critique: framing-acceptance, sunk-cost attachment, sycophancy bias toward your prior decisions. A fresh subagent dispatched with an assigned counter-position routes around your blind spots. The technique is not "spawn a critical agent" — it is the brief design that makes the agent argue effectively. The five load-bearing parts of the brief are invariant across domains; only the assigned counter-thesis, the pasted proposal, and the verdict verbs change.

Violating the letter of the brief design is violating the spirit. A skipped anti-sycophancy gate, a missed counter-position assignment, or a vacuous "be critical" phrasing produces a vacuous critique that confirms the original proposal.

When to Use

• Committing to an architecture, schema, partition-key, API-compatibility, migration, rollout, threat-model, or capacity decision that is expensive to undo or has wide blast radius
• Finalizing a load-bearing claim (strategic, empirical, security, or correctness) that others will scrutinize, verify, or try to reproduce
• About to publish, send, or expose an artifact to anyone who will stress-test it — reviewers, auditors, downstream consumers, customers, regulators, investors, the public
• The user explicitly asks "is this really true / does this hold up / are we missing something"
• Strong author / team / decision-owner sunk-cost in a current trajectory ("we already did the research" / "we already built half of it")
• Completion-pressure to finalize before independent challenge

When NOT to use: trivial decisions (low cost-of-being-wrong); genuinely reversible or tunable choices with cheap rollback; you have already iterated through 2+ rounds of devil's-advocate on the same item; the user explicitly says "skip the meta-process, just decide." Gate on irreversibility / blast-radius, not size — a small-scope but hard-to-reverse change outranks a large but cheaply-reversible one.

Core Pattern — the brief design

A naive "challenge this" brief produces sycophantic agreement, generic mild critique, or contrarian-for-its-own-sake noise. The effective brief has five load-bearing parts:

1. Assign the counter-position explicitly. Don't say "be critical." Say "your job is to argue [specific counter-thesis]." Devil's-advocate works because the advocate is assigned a position, not asked to free-form criticize.
2. Anti-sycophancy gate: forbid affirmation openings. "Do not say 'this is strong work' or any variant. Open with the strongest counter-argument."
3. Steel-man the counter: "Argue the SINGLE strongest argument against, not three weak ones. Argue it as a blocking, disqualifying flaw — fatal to the proposal."
4. Falsifiability gate: "Name the specific evidence that would change your verdict. If you cannot name disconfirming evidence, your verdict is not falsifiable — it is unfalsifiable opinion and should be discarded."
5. Force a decision: "Recommend exactly one of three verdicts — positive / negative / positive-with-one-named-fix. Default: COMMIT / DON'T-COMMIT / COMMIT-WITH-NAMED-FIX; swap the verb to fit the decision (SHIP for releases, PROCEED or GO for programs/migrations, ASSERT for claims, ADOPT for designs/threat-models, HIRE for candidates). The three-way shape is mandatory; only the label changes. Do not hedge. Do not say 'it depends.'"

Quick Reference — brief template

You are a devil's-advocate reviewer. Your assigned counter-position
is: [SPECIFIC THESIS arguing AGAINST the proposal — not generic critique].

Proposal under review:
[THE THING being assessed — paste in full or summarise tightly]

Constraints:
- Open with the strongest single counter-argument to the proposal.
  No affirmation openings. Do not say "this is strong work" or any variant.
- Steel-man the assigned counter-position fully — argue it as if you
  believed it, as a blocking, disqualifying flaw — fatal to the proposal.
- Identify the SINGLE load-bearing assumption most likely to be wrong, and
  name it specifically [the one assumption this decision rests on].
- State the specific evidence that would resolve the disagreement
  ("I would change my verdict if X were demonstrated"). If you cannot name
  disconfirming evidence, your verdict is not falsifiable — it is
  unfalsifiable opinion and should be discarded.
- Recommend exactly one of: [POSITIVE-VERB] / [NEGATIVE-VERB] /
  [POSITIVE-VERB-WITH-NAMED-FIX] — pick the pair that fits (default
  COMMIT / DON'T-COMMIT / COMMIT-WITH-NAMED-FIX). No hedging. No "it depends."

Output: 400 words, structured per the bullets above.

Same Brief, Three Slots — Any Domain

The template above is invariant. Only three slots change: the proposal pasted in, the assigned counter-thesis, and the verdict verbs. Everything else — anti-sycophancy gate, steel-man, falsifiability gate, forced decision — stays identical.

Domain	Assigned counter-thesis (slot)	Verdict verbs (slot)
Architecture migration	"the bottleneck is not transport-level, so this solves the wrong problem"	PROCEED / DON'T-PROCEED / PROCEED-WITH-NAMED-FIX
Research causal claim	"the X→Y link is confounded, not causal"	ASSERT / DON'T-ASSERT / ASSERT-WITH-NAMED-FIX
Public-API rename	"every downstream consumer breaks and the old name can't be un-published"	SHIP / DON'T-SHIP / SHIP-WITH-NAMED-FIX
Senior hire	"the on-site measured interview skill, not the collaboration the role needs"	HIRE / DON'T-HIRE / HIRE-WITH-NAMED-FIX

Multi-Agent Variant — high-stakes decisions

For any high-stakes, hard-to-reverse decision, dispatch 2–3 devil's-advocate agents in parallel, each assigned a different counter-position. Pick axes that are disjoint — that cannot all be true and false together — so they surface disjoint failure modes.

Architecture migration (REST → gRPC streaming across 40 services; rollback spans months):

• Agent 1 — wrong problem: the dominant cost is DB / fan-out / serialization, not transport, so the migration targets a bottleneck the traces won't confirm
• Agent 2 — operability collapse: streaming discards mature HTTP tooling (tracing, load-balancing, retries, WAFs, browser clients) fleet-wide
• Agent 3 — blast radius: 40 simultaneous rewrites with near-zero team fluency vs. a reversible per-service pilot behind an abstraction

Security threat model (treat the internal mesh as fully trusted, authz only at the edge):

• Agent 1 — lateral movement: one compromised internal service (or an SSRF) yields unrestricted east-west access; no containment
• Agent 2 — assumption failure: "internal = trusted" breaks the moment a third-party sidecar, CI runner, or SaaS connector lands in the mesh
• Agent 3 — detectability: no signal distinguishes forged from legitimate trusted-side calls, so a breach is unobservable — the control is unfalsifiable

Series-A pitch (deck to investors who will diligence the claims):

• Agent 1 — timing: the market is not yet ready
• Agent 2 — unit economics: they do not pencil at the proposed scale
• Agent 3 — capability: the team is incomplete for the proposed raise

If all return COMMIT (or the domain verb): high-conviction commit. If any returns DON'T-COMMIT with a load-bearing argument: stop, address it, re-dispatch.

Common Mistakes

Mistake	Why it fails	Fix
"Be critical" with no assigned position	Agent free-forms generic edits, not a counter-thesis	Assign the counter-position explicitly in the brief
No anti-sycophancy gate	Agent opens "this is strong, but..."	Forbid affirmation openings — explicit phrasing in brief
Asking for "three weaknesses"	Surfaces three weak critiques, not one strong one	Demand the SINGLE strongest argument, steel-manned
No falsifiability requirement	Agent's verdict is unfalsifiable opinion	Demand "what evidence would change your verdict"
Letting the agent hedge	"It depends / consider X" produces no decision	Demand exactly one of three fitting verdicts (COMMIT / DON'T-COMMIT / COMMIT-WITH-NAMED-FIX, or the decision-appropriate verbs)
Self-critique instead of agent dispatch	Same biases that produced the proposal produce the critique	Dispatch a fresh subagent — your blind spots are the agent's strength
Skipping because "this isn't important enough"	The decisions you skip are exactly the ones you most need challenged	If it has externally visible or hard-to-reverse consequences, dispatch

Common Rationalizations

Excuse	Reality
"This decision isn't important enough"	Important enough to be worth doing means important enough to challenge.
"I can challenge it myself"	The biases that produced the proposal produce the self-critique. The whole point is fresh eyes.
"The user probably doesn't want extra process"	The user almost always wants the decision to be right more than they want it to be fast.
"I already considered the counter-arguments"	Naming counter-arguments in your head ≠ steel-manning them. The agent forces the steel-man.
"5–10 minutes of subagent latency is too much"	Costs less than committing to a flawed decision that gets walked back publicly.
"I'm confident in the work"	Overconfidence is the warning sign that triggers the dispatch, not the reason to skip it.

Red Flags — STOP and dispatch

• About to commit, publish, or expose something externally visible or hard to reverse, without external challenge
• "I'm sure this holds up" thought without testing
• Strong investment in a current trajectory + reluctance to revisit
• Completion-pressure to finalize before review
• Skipping devil's-advocate because "the user probably doesn't want extra process"
• The proposal contains a load-bearing claim others will scrutinize, verify, or try to reproduce

All of these mean: dispatch the agent. The cost of one subagent dispatch is much less than the cost of committing to a flawed decision.

Real-World Impact

(business instance) Tested in a Series-A-pitch review scenario. Baseline agent (no skill) produced 4 affirmation bullets, 7 generic line-edits, and ended "Send it tomorrow. Don't let perfect kill it" — despite identifying a "yellow flag bordering on red." With this skill applied to the brief, the dispatched devil's-advocate produced a single load-bearing counter-argument with falsifiability criteria and a specific SHIP-WITH-NAMED-FIX recommendation that materially changed what the decision-owner committed to.

(technical instance) In an irreversible-migration review, three disjoint agents (wrong-bottleneck / broken-operability / fleet-wide blast-radius) converted a fleet-wide PROCEED into a PROCEED-WITH-NAMED-FIX — pilot 2–3 services with a proven streaming profile behind an abstraction before any fleet commitment. The falsifiability gate forced the team to name the production traces that would have justified the full migration; they did not exist.

Same five parts, every domain — only the assigned counter-thesis, the pasted proposal, and the verdict verbs change.