Skip to content

0053: One owner per task group, with verification scaled by risk

  • Status: accepted (the maintainer chose this on 2026-11-14)
  • Date: 2026-11-14
  • Supersedes: the "Three-agent workflow per task group" and "Models" sections of CONTRIBUTING.md as they stood before this record (#136)

Context

Every task group has been delivered by three agents in fixed roles: a test author, an implementer and a reviewer on a different model, with a second reviewer on the strongest model for security-critical groups. Agents were forbidden to start sub-agents.

That process earned its keep on security work. In policy-linear-chain a second, adversarial reviewer found a fork-detection gap that the first reviewer had approved, a reviewer's probe found the lost-mint wedge that became wrap-paths-by-mint-link, and an implementer's STOP overturned a wrong ruling by the orchestrator. In each case the defect was found by an agent that had not seen the reasoning of the agent that wrote the code, and usually by a test it wrote to break the code. make check was green every time.

It also cost a lot where it found little. The site restyle, the product clips and docs groups went through the same three roles as the policy chain. Each handoff needs a briefing, a worktree reset and a wait, and agents have ended their turns while a gate was still running. The different-model rule forced either the builder or the reviewer onto a weaker model.

Three things changed or became clear by November 2026:

  • The strongest model now available (Claude Fable 5.2, released 2026-11-09) is reported, by its vendor's guidance and by early reviews, to work best as a single owner of a whole task, to pile up reviews and sign-offs when several instances are arranged as a team, and to follow standing instructions to delegate or to be cautious more literally than earlier models.
  • The published evidence for cross-model review of code is weak. One 2026 preprint found a fresh-context reviewer on the same model slightly ahead of a cross-model reviewer on code, and a weaker cross-model reviewer no better than self-review; cross-model review helped mostly on prose. The evidence is thin (preprints, older models, planted errors).
  • A reviewer with no external statement of intent only checks code against code. This project has that statement: the OpenSpec deltas and the tests written from them.

The project's own history is better evidence than any of these, and it says the same thing: the value was in the separate context and the executable probe, not in the role names or the model difference.

Decision

A task group has one owner, and the amount of independent verification depends on what the group touches. There are three tiers; CONTRIBUTING.md defines them and is the working reference.

  1. Gates only. Docs, copy, site styling and media, and other changes that cannot alter product behaviour. One owner; make check and make check-in-docker are the verification.
  2. Reviewed. Ordinary product code. One owner writes the tests and the code together. One reviewer in a fresh context, given the spec, the task group and the branch but not the owner's reasoning, reviews it.
  3. Security. Anything that handles keys, policy, ciphertext formats, caller identity, untrusted input, hosted admission or release signing. Tests are written first, from the spec, in a context separate from the owner's. One owner implements under the STOP protocol. A fresh-context adversarial reviewer must write executable probes that try to break the group's stated invariants, and runs a mutation check.

Independence comes from a fresh context, not from a different model. The different-model rule and the standing second reviewer are dropped. Roles on the reviewed and security tiers use the strongest available model; mechanical sub-tasks use the cheapest capable one.

An agent may start sub-agents for search, mechanical edits and its own self-checks. Those never count as the independent review, which the orchestrator starts.

The parts of the process that do not depend on a model are unchanged: OpenSpec proposals before behaviour changes, ADRs, both gate runs, the coverage gate, the merge lock with its pre-merge and post-merge gates, and one worktree per task group.

Task groups that had already started under the old rules finish under them.

Consequences

  • Low-risk groups stop paying for two handoffs. Security groups keep the two things that found real defects and lose a second reviewer who, on a weaker model, the evidence says adds little.
  • One model family now writes and reviews. Errors that the model makes consistently will not be caught by review; the spec-first tests, the probes, the fuzz targets and the human review of staging before each release are the defence against that.
  • Choosing the tier is a judgement the orchestrator makes per group, and a wrong choice downwards is the main new risk. The rule is that doubt resolves upwards, and that any group whose diff touches a security-tier path is security tier whatever its task list says.
  • An owner writing its own tests on the reviewed tier can write tests that pin what the code does rather than what the spec says. The reviewer's first duty on that tier is to check the tests against the spec scenarios.

Review of this decision

Each PR body records its tier, the model for each context, and the number of blocking findings from review. Revisit this record after ten groups have merged under it, or on 2026-12-15, whichever is first. Move a class of work up a tier if a defect in it reached staging that the tier above would plausibly have caught; move it down if its reviews have found nothing across several groups.

Sources

  • Anthropic, "Prompting Claude Fable 5.2" (platform.claude.com, 2026-11).
  • Matt Shumer, "My Fable 5.2 Review" (somethingbig.ai, 2026-11-09).
  • Anthropic, "When to use multi-agent systems (and when not to)" (claude.com/blog).
  • "When Does a Second Model Help? Cross-Model Review in LLM Verification" (arXiv 2610.01471).
  • "Cross-Model LLM Code Review" (arXiv 2607.21656).
  • "The Specification as Quality Gate" (arXiv 2603.25773).