Council pattern

The core nax workflow shape is a council of agents: run selected providers independently, pass their findings into a cross-review step, then synthesize only the claims with enough evidence or useful disagreement.

This is not majority voting. The goal is to preserve independent judgment before the agents influence each other, then make disagreements inspectable before a person acts on the result.

Step through a full council run below — or press play.

1/7 · The flow file

A workflow is a flow config plus Markdown prompts — diffable, reviewable, repeatable. This flow has three steps; nax runs them in order.

Shape

The first pass should ask each agent to inspect the same task without seeing the other answers. Later steps can include earlier results so agents can confirm, reject, duplicate, or challenge specific findings.

Why it works

Different models have different strengths and failure modes.
Agreement is more useful when it happens after independent first passes.
Disagreement is useful when it includes concrete evidence.
The final synthesis can filter plausible but unsupported claims.
Saved artifacts make the run auditable instead of reducing it to one final blob.

Common workflows

Use the council pattern for work that benefits from structured disagreement before action:

Code review: inspect a diff, cross-check findings, and summarize the issues a human should review.
Security review: focus on auth, billing, webhooks, data exposure, dependencies, and deployment configuration.
Bug triage: classify reports, propose likely causes, rank severity, and recommend diagnostics.
Feature planning: compare model proposals against docs, user pain, analytics notes, and repo constraints.
Bake-offs: ask agents for competing implementation plans, then compare complexity, risk, and expected quality.
PR repair: analyze a failing PR, build a fix plan, then let one agent implement after review.

Runtime agent and model choice

Provider, model, and effort choice can all be part of the run:


nax run review \
  --agents claude,codex \
  --models claude=claude-opus-4-8 \
  --efforts claude=high
nax run review --step-agents cross-review=gemini --step-agents synthesize=codex

Bundled councils default to Claude, Gemini, and Codex. OpenCode is a selectable fourth member — add it with --agents claude,gemini,codex,opencode (or per step) when you want a wider spread of models.

Use the full council when a task is high risk or ambiguous. Use a smaller agent set for cheap triage, targeted follow-up, or quick checks before spending more tokens.

Human authority

nax can do exploration, synthesis, and first-pass implementation, but the workflow should still make approval explicit when the result can change code or affect production decisions. Use a human review gate or route the final artifact into a PR so a person owns the merge decision.