The Automaton Auditor: Multi-Agent Code Auditing Swarm
A LangGraph swarm that grades repositories against a structured rubric. Detective, Prosecutor, Defense, Tech Lead, and Chief Justice nodes produce a deterministic verdict.
10 Academy TRP · 1-week briefVerdict flow
Highlights
- ▸Adversarial multi-agent pattern: Prosecutor argues issues, Defense counters, Tech Lead adjudicates, Chief Justice returns a verdict.
- ▸Detective stage runs deterministic forensic analysis on the repo before any LLM reasoning happens.
- ▸Structured rubric produces reproducible verdicts and a concrete remediation plan.
The idea
Ordinary LLM code reviewers agree with themselves too much. If the model likes a pattern, it will keep liking that pattern across every file, and its “review” is really a repeated preference.
The Auditor forces disagreement into the loop. A Prosecutor node argues the negative case against every finding. A Defense node argues the positive. A Tech Lead mediates, and a Chief Justice returns the final verdict: deterministic, cited, and reproducible.
Pipeline
- Detective: deterministic forensic analysis over the repo (AST, dependency graph, file metadata). No LLM.
- Prosecutor: argues issues, grounded in Detective’s evidence.
- Defense: argues against, grounded in the same evidence.
- Tech Lead: weighs the arguments.
- Chief Justice: signs off on the verdict and produces a remediation plan.
Why it works
Structured disagreement produces more calibrated verdicts than a single reviewer prompt. The Detective stage keeps the reasoning grounded in actual code artifacts, not vibes.