Multi-agent adversarial verification with convergence loop. Two independent review agents must both pass before output ships.
Multi-agent adversarial verification framework. Make a list, check it twice. If it's naughty, fix it until it's nice.
The core insight: a single agent reviewing its own output shares the same biases, knowledge gaps, and systematic errors that produced the output. Two independent reviewers with no shared context break this failure mode.
Invoke this skill when:
Do NOT use for internal drafts, exploratory research, or tasks with deterministic verification (use build/test/lint pipelines for those).
┌─────────────┐
│ GENERATOR │ Phase 1: Make a List
│ (Agent A) │ Produce the deliverable
└──────┬───────┘
│ output
▼
┌──────────────────────────────┐
│ DUAL INDEPENDENT REVIEW │ Phase 2: Check It Twice
│ │
│ ┌───────────┐ ┌───────────┐ │ Two agents, same rubric,
│ │ Reviewer B │ │ Reviewer C │ │ no shared context
│ └─────┬─────┘ └─────┬─────┘ │
│ │ │ │
└────────┼──────────────┼────────┘
│ │
▼ ▼
┌──────────────────────────────┐
│ VERDICT GATE │ Phase 3: Naughty or Nice
│ │
│ B passes AND C passes → NICE │ Both must pass.
│ Otherwise → NAUGHTY │ No exceptions.
└──────┬──────────────┬─────────┘
│ │
NICE NAUGHTY
│ │
▼ ▼
[ SHIP ] ┌─────────────┐
│ FIX CYCLE │ Phase 4: Fix Until Nice
│ │
│ iteration++ │ Collect all flags.
│ if i > MAX: │ Fix all issues.
│ escalate │ Re-run both reviewers.
│ else: │ Loop until convergence.
│ goto Ph.2 │
└──────────────┘
Execute the primary task. No changes to your normal generation workflow. Santa Method is a post-generation verification layer, not a generation strategy.
# The generator runs as normal
output = generate(task_spec)
Spawn two review agents in parallel. Critical invariants:
REVIEWER_PROMPT = """
You are an independent quality reviewer. You have NOT seen any other review of this output.
## Task Specification
{task_spec}
## Output Under Review
{output}
## Evaluation Rubric
{rubric}
## Instructions
Evaluate the output against EACH rubric criterion. For each:
- PASS: criterion fully met, no issues
- FAIL: specific issue found (cite the exact problem)
Return your assessment as structured JSON:
{
"verdict": "PASS" | "FAIL",
"checks": [
{"criterion": "...", "result": "PASS|FAIL", "detail": "..."}
],
"critical_issues": ["..."], // blockers that must be fixed
"suggestions": ["..."] // non-blocking improvements
}
Be rigorous. Your job is to find problems, not to approve.
"""
# Spawn reviewers in parallel (Claude Code subagents)
review_b = Agent(prompt=REVIEWER_PROMPT.format(...), description="Santa Reviewer B")
review_c = Agent(prompt=REVIEWER_PROMPT.format(...), description="Santa Reviewer C")
# Both run concurrently — neither sees the other
The rubric is the most important input. Vague rubrics produce vague reviews. Every criterion must have an objective pass/fail condition.
| Criterion | Pass Condition | Failure Signal |
|---|---|---|
| Factual accuracy | All claims verifiable against source material or common knowledge | Invented statistics, wrong version numbers, nonexistent APIs |
| Hallucination-free | No fabricated entities, quotes, URLs, or references | Links to pages that don't exist, attributed quotes with no source |
| Completeness | Every requirement in the spec is addressed | Missing sections, skipped edge cases, incomplete coverage |
| Compliance | Passes all project-specific constraints | Banned terms used, tone violations, regulatory non-compliance |
| Internal consistency | No contradictions within the output | Section A says X, section B says not-X |
| Technical correctness | Code compiles/runs, algorithms are sound | Syntax errors, logic bugs, wrong complexity claims |
Content/Marketing:
Code:
any leaks, proper null handling)Compliance-Sensitive (regulated, legal, financial):
def santa_verdict(review_b, review_c):
"""Both reviewers must pass. No partial credit."""
if review_b.verdict == "PASS" and review_c.verdict == "PASS":
return "NICE" # Ship it
# Merge flags from both reviewers, deduplicate
all_issues = dedupe(review_b.critical_issues + review_c.critical_issues)
all_suggestions = dedupe(review_b.suggestions + review_c.suggestions)
return "NAUGHTY", all_issues, all_suggestions
Why both must pass: if only one reviewer catches an issue, that issue is real. The other reviewer's blind spot is exactly the failure mode Santa Method exists to eliminate.
MAX_ITERATIONS = 3
for iteration in range(MAX_ITERATIONS):
verdict, issues, suggestions = santa_verdict(review_b, review_c)
if verdict == "NICE":
log_santa_result(output, iteration, "passed")
return ship(output)
# Fix all critical issues (suggestions are optional)
output = fix_agent.execute(
output=output,
issues=issues,
instruction="Fix ONLY the flagged issues. Do not refactor or add unrequested changes."
)
# Re-run BOTH reviewers on fixed output (fresh agents, no memory of previous round)
review_b = Agent(prompt=REVIEWER_PROMPT.format(output=output, ...))
review_c = Agent(prompt=REVIEWER_PROMPT.format(output=output, ...))
# Exhausted iterations — escalate
log_santa_result(output, MAX_ITERATIONS, "escalated")
escalate_to_human(output, issues)
Critical: each review round uses fresh agents. Reviewers must not carry memory from previous rounds, as prior context creates anchoring bias.
Subagents provide true context isolation. Each reviewer is a separate process with no shared state.
# In a Claude Code session, use the Agent tool to spawn reviewers
# Both agents run in parallel for speed
# Pseudocode for Agent tool invocation
reviewer_b = Agent(
description="Santa Review B",
prompt=f"Review this output for quality...\n\nRUBRIC:\n{rubric}\n\nOUTPUT:\n{output}"
)
reviewer_c = Agent(
description="Santa Review C",
prompt=f"Review this output for quality...\n\nRUBRIC:\n{rubric}\n\nOUTPUT:\n{output}"
)
When subagents aren't available, simulate isolation with explicit context resets:
The subagent pattern is strictly superior — inline simulation risks context bleed between reviewers.
For large batches (100+ items), full Santa on every item is cost-prohibitive. Use stratified sampling:
import random
def santa_batch(items, rubric, sample_rate=0.15):
sample = random.sample(items, max(5, int(len(items) * sample_rate)))
for item in sample:
result = santa_full(item, rubric)
if result.verdict == "NAUGHTY":
pattern = classify_failure(result.issues)
items = batch_fix(items, pattern) # Fix all items matching pattern
return santa_batch(items, rubric) # Re-sample
return items # Clean sample → ship batch
| Failure Mode | Symptom | Mitigation |
|---|---|---|
| Infinite loop | Reviewers keep finding new issues after fixes | Max iteration cap (3). Escalate. |
| Rubber stamping | Both reviewers pass everything | Adversarial prompt: "Your job is to find problems, not approve." |
| Subjective drift | Reviewers flag style preferences, not errors | Tight rubric with objective pass/fail criteria only |
| Fix regression | Fixing issue A introduces issue B | Fresh reviewers each round catch regressions |
| Reviewer agreement bias | Both reviewers miss the same thing | Mitigated by independence, not eliminated. For critical output, add a third reviewer or human spot-check. |
| Cost explosion | Too many iterations on large outputs | Batch sampling pattern. Budget caps per verification cycle. |
| Skill | Relationship |
|---|---|
| Verification Loop | Use for deterministic checks (build, lint, test). Santa for semantic checks (accuracy, hallucinations). Run verification-loop first, Santa second. |
| Eval Harness | Santa Method results feed eval metrics. Track pass@k across Santa runs to measure generator quality over time. |
| Continuous Learning v2 | Santa findings become instincts. Repeated failures on the same criterion → learned behavior to avoid the pattern. |
| Strategic Compact | Run Santa BEFORE compacting. Don't lose review context mid-verification. |
Track these to measure Santa Method effectiveness:
Santa Method costs approximately 2-3x the token cost of generation alone per verification cycle. For most high-stakes output, this is a bargain:
Cost of Santa = (generation tokens) + 2×(review tokens per round) × (avg rounds)
Cost of NOT Santa = (reputation damage) + (correction effort) + (trust erosion)
For batch operations, the sampling pattern reduces cost to ~15-20% of full verification while catching >90% of systematic issues.