Redteam

Redteam

A one-shot subagent that challenges work at a milestone. The main agent decides when to spawn it; the subagent reads the relevant artifacts and returns either "nothing substantive to challenge" or a list of concrete findings.

When to spawn

Run a redteam at milestones, not continuously:

• Goal locked. Right after init_research(goal) — is the goal well-posed, are there obvious framings being missed?
• Major established result. A claim has been verified and is about to anchor downstream work. Worth one challenge before it becomes load-bearing.
• Before write-up. The conclusion is drafted and you're about to present it. Last pass before it leaves your hands.
• When the user asks. "Can you push back on this?" → spawn a redteam.

Do not run redteam continuously. An agent told to find mistakes will find phantom ones if asked too often. Milestone cadence preserves signal.

How to spawn

Task(
  subagent_type="redteam",
  prompt=(
    "Milestone: <what was just reached>\n"
    "Artifact to challenge: <file path or claim ID or excerpt>\n"
    "Optional axis to focus on: <e.g. 'unstated assumptions', 'alternative methods'>\n"
    "\n"
    "Return your verdict and any findings per the output format in your "
    "system prompt. Empty findings is the right answer if the work is sound."
  )
)

The subagent reads the artifact, decides whether there's anything substantive to challenge, and returns a structured report. It does not write to research state or modify any files.

Reading the result

The subagent returns one of four verdicts:

• "nothing substantive to challenge" — the work is sound. Move on. Logging this in workspace/research/redteam.log is optional (one line if you want a record).
• "minor presentational notes only" — clarity could improve but the conclusion holds. Apply if it helps the next reader.
• "issues worth addressing" — concrete findings that should be resolved before the work is final.
• "concerns that block the conclusion" — the result is in question. Re-open the analysis.

If the verdict is "issues worth addressing" or higher, treat each must-address finding as a new claim to investigate — verify, refute, or reframe.

Manual invocation

/redteam runs a single redteam pass against whatever the main agent thinks is the current milestone. The user can also pass an explicit target:

/redteam workspace/research/established.md C3
/redteam "the conclusion that the period scales as L/g"

Hard rules

• Milestone-only by default. Don't spawn redteam mid-derivation.
• Trust empty verdicts. If the subagent says "nothing substantive," don't ask again with a leading prompt to extract findings.
• One axis per spawn. If you want to challenge both methodology and assumptions, spawn twice with different focus axes — keeps each pass calibrated.
• The subagent is read-only. Do not ask it to write to research state or files.