Docs / Workflow

The AI red team pass

The red team pass is an adversarial second opinion. Where the main generation writes the analysis or the canvas, the red team’s only job is to attack it: to find where it is wrong, unfounded, or quietly papering over a gap, and to say so in front of the reviewer.

The red team pass is a Business-plan capability, and it is checked server-side. On other plans the option isn’t offered, and generation runs exactly as it otherwise would.

It exists because an AI marking its own homework is worth very little. A model asked to produce a document and then asked whether that document is good will nearly always say yes. Giving a separate pass the explicit job of finding fault, with no stake in defending the draft, gets you a genuinely critical read instead of a rubber stamp.

Turning it on

The red team pass is opt-in per stage, so you can spend it where the risk is. There are two places to switch it on, one for each artifact it can critique:

On the analysis
Tick “AI red team review on submit” when you submit a draft for analysis, either in the feature editor or in the draft action row. The pass runs on the analysis document once it is generated.
On the canvases
Tick the same option when you approve the analysis. The pass runs on each service’s REASONS canvas as it is generated, critiquing each one separately.

Neither choice is sticky and neither is all-or-nothing: you can red team a risky analysis and skip the canvases, or wave through a routine analysis and scrutinize the build plans. Both are left off by default.

What it actually checks

The red team reads the finished artifact as a critic rather than an author, hunting for the failure modes a confident-sounding document hides best:

  • Claims the document asserts but cannot support.
  • Risks, edge cases, and failure modes it should have raised and didn’t.
  • Acceptance criteria the plan doesn’t genuinely cover, however much it appears to.
  • Reasoning that is internally inconsistent, or that quietly contradicts itself between sections.

It doesn’t judge the document only on its own terms. Where a service has a repo connected, the red team can go and read your actual code, so it can catch a plan that describes concepts or constraints that don’t exist. When it critiques a canvas, it also reads the approved analysis the canvas was built from, so it can call out a build plan that has drifted from the ask it was supposed to serve.

Findings arrive as comments

The pass posts its critique as a comment on the artifact it reviewed, authored by the AI Red Team and carrying its own badge, sitting unresolved in the collaboration sidebar exactly where the reviewer is already looking. It behaves like any other comment: you can reply to it, argue with it, and resolve it.

That placement is the point. The critique lands next to the document it is about, at the moment the human gate opens, so the reviewer reads the plan and its sharpest objection side by side rather than having to think of the objection themselves.

The red team advises; it never decides. It cannot fail a generation, block a feature, or send anything back. It leaves a finding, and a human reviewer weighs it like any other comment and still makes the call at the gate.

When to reach for it

It is a second AI call, so it is worth spending deliberately rather than leaving on everywhere:

  • A feature touching a sensitive or regulated area, where a missed risk is expensive.
  • A thin or ambiguous definition, where the analysis had to make judgement calls to fill the gaps.
  • A plan whose reviewer isn’t deep in that service, and who would benefit from a second read pointing at the soft spots.
Why this exists
The weakest moment in an AI-assisted workflow is the one where a document reads well enough that nobody argues with it. A fluent plan is persuasive whether or not it is correct, and a reviewer facing a polished artifact with no opposition tends to approve it. The red team’s job is to guarantee there is opposition: a critical voice in the room every time, with an explicit mandate to find fault and no interest in defending the draft. It makes the human gate a real decision rather than a formality, which is the entire reason the gate exists.