Home / Workshop, capstone and exam
Advanced capstone, the multi agent adversarial review
An orchestrator that distributes narrow roles to several reviewers in parallel, a step that blocks incomplete outputs, then a synthesis that surfaces conflicts rather than smoothing them over, produce a more reliable review than a single reviewer.
A single reviewer examining a deliverable often falls into the same blind spot as whoever wrote the code, because they share the same starting reasoning. Distributing the review across several subagents, each given a narrow role, corrects this bias, provided three things hold together: the roles stay distinct, no incomplete output passes on to the next step, and the final synthesis surfaces disagreements instead of smoothing them into a single conclusion.
Narrow roles, not clones
Four subagents launched with the same general instruction produce four variations on the same opinion, because they start from the same scope. A narrow role, by contrast, bounds what each reviewer looks at: one checks only functional correctness, another only the security surface, a third only the simplicity of the code, a fourth only edge case coverage. Each role is written in a few lines, with an explicit scope and a ban on commenting on what falls outside it.
rôle : relecteur-sécurité
périmètre : entrées utilisateur, permissions, secrets
interdit : commenter le style ou la performance
sortie attendue : liste de constats, chacun avec fichier, ligne, scénario d'échecClaude Code subagents run in parallel, each in its own isolated context, with its own tools and permissions. This isolation stops one reviewer from being influenced by another's reasoning before delivering its own opinion, which is exactly what a narrow role is meant to achieve.
The synthesis must never smooth things over silently
The fragile point in this setup is not the parallel review itself, it is the step that follows it. Nothing in the product automatically surfaces a conflict between two reviewers: it is the synthesis agent, whether that is the main conversation or a coordinator subagent written for the purpose, that must be explicitly instructed to flag a disagreement rather than choosing alone which of the two opinions to keep. Without this instruction set in advance, the synthesis averages the opinions in silence, and the disagreement disappears from the final report as though it had never existed.
The same discipline applies to blocking incomplete outputs. Before launching the synthesis, check that each reviewer has actually delivered a finding for the whole of its scope, not just for the portion of the file it had time to read. A role that returns a partial opinion has the same format as a role that finished its work, nothing visually distinguishes them, so nothing distinguishes them for the synthesis either unless you have planned for it yourself.
This project directly extends the choice made in the three project capstone: the adversarial review is not a separate fourth project, it is a more demanding way to check any of the three before presenting it as finished.
The path of the adversarial review
To deliver their capstone, a practitioner writes four distinct reviewer roles, each with its own scope, launches the four subagents in parallel on the same deliverable, and waits for all four to deliver their findings before opening the synthesis report.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: This starting point establishes that four independent and distinct opinions will exist on the same deliverable, each bounded to its own scope.
What this does not establish: It does not establish that the synthesis report, once opened, will surface disagreements between these four opinions rather than averaging them in silence.
The three most common miscalibrations
- Too broad This situation establishes that the practitioner's capstone will succeed, since four distinct reviewers have already examined the deliverable.
- Too narrow This situation establishes nothing at all, since the number of roles defined never changes the quality of a review.
- Beside the point This situation establishes that the deliverable submitted for review already meets the acceptance criteria set for the project.
- A review entrusted to a single agent systematically inherits its own blind spot, since there is no one to contradict it.
- A reviewer role is defined as much by what it is banned from commenting on as by what it must check.
- Context isolation between subagents stops one reviewer from copying another's opinion before delivering its own.
- No mechanism in the product surfaces a disagreement in place of the practitioner, this instruction must be written explicitly into the synthesis role.
- A partial finding that has not covered the whole of the stated scope must be treated as an incomplete output, not as a valid opinion.
Write a narrow role for each of your four reviewers, explicitly banning them from commenting on anything outside their scope, then add a line to your synthesis instruction that requires listing every disagreement between reviewers rather than deciding it alone.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, subagents in Claude Code consultée le 2026-09-02