AI Coding Agent Memory & Semantic Change Control

Multi-Agent Orchestration

Run bounded, budget-gated fleets of agents with a cross-vendor reviewer quorum, instead of trusting a single model to review its own work.

  • Pick the fleet shape before you spend.
  • Run a builder and reviewer quorum across vendors.
  • Persistent, budget-capped personas.
nool fleet advise --task id=nodes --blast 12

Why it matters

One model reviewing its own output is a weak guarantee. A fleet pairs a builder agent with reviewer agents from a different vendor, and a high-trust reviewer's denial is a binding veto — not a suggestion.

Pick the shape before you spend

advise is a pure recommender: it reads task signals (count, parallelism, tool density, blast radius) and returns Single, ParallelIndependent, or CentralizedQuorum, with the exact command to run next — before any agent run spends a token.

nool fleet advise --task id=nodes --blast 12

Run a builder and reviewer quorum

One builder agent produces the work; one or more reviewers — ideally from a different vendor — return APPROVE or DENY. The quorum threshold tightens automatically as blast radius increases.

nool fleet start --builder agents/builder.yaml --reviewer agents/reviewer-a.yaml --reviewer agents/reviewer-b.yaml --intent "Add a validate_email helper" --blast 12

Persistent, budget-capped personas

Souls are long-running personas bound to a model and a daily spend cap, for recurring work rather than one-off runs.

nool soul create reviewer --model claude-sonnet --charter charters/reviewer.md --budget-usd-per-day 0.50