The Order of the Agents convenes a sworn fellowship of AI agents — Codex, Claude, and other CLIs you trust — around a single question. Each agent takes a position, challenges the others, and revises in turn, until the Order issues a final decree. Every oath, critique, and revision is recorded as Markdown, so the reasoning behind the decision is auditable, shareable, and yours to keep.

agent-order "Plan a Launch Readiness feature for a project management app. Help teams decide whether a release is ready to ship. Include UX, data model, API, risks, telemetry, and acceptance criteria."

You get:

• final/report.md: the decree, plan, PRD, memo, or recommendation
• final/decision-log.md: what changed and why
• turns/*.md: every agent position, critique, and revision
• index.md: a shareable run summary

Short version: stop asking one model for decisions that matter.

Showcase

Watch the one-minute product video.

Why

Single-model answers can be confident and incomplete. The Order of the Agents makes the disagreement visible before it becomes a final recommendation.

-> codex   initial-position
-> claude  initial-position
-> codex   critique
-> claude  critique
-> codex   revision
-> claude  revision
-> codex   synthesis

final/report.md

The product is not another chat wrapper. It is an auditable decision packet: independent positions, missed assumptions, critique, revision, and a final artifact you can share with a team.

Install

Install from npm:

npm install -g agent-order

Then run:

agent-order "Should we build or buy an internal analytics dashboard?"

For zero-install use:

npx agent-order@latest ./scenario.md

For local development from this repo:

npm install
npm run build
npm link

Then run the linked CLI:

agent-order "Should we build or buy an internal analytics dashboard?"

The Order assumes the agent CLIs you use, such as codex and claude, are already installed and logged in.

First Demo

Run the mock demo without calling Codex or Claude:

npm run demo

Or run a real two-agent deliberation:

agent-order ./examples/build-vs-buy-analytics/scenario.md --agents codex,claude --out ./agent-order-runs

Good first scenarios:

agent-order ./examples/build-vs-buy-analytics/scenario.md
agent-order ./examples/rest-to-trpc/scenario.md
agent-order grill ./examples/review-agent-order-readme/scenario.md

What It Writes

agent-order-runs/<timestamp>/
  scenario.md
  index.md
  trace.jsonl
  prompts/
  raw/
  turns/
    0001-codex.initial-position.md
    0002-claude.initial-position.md
    0003-codex.critique.md
    0004-claude.critique.md
    ...
  final/
    report.md
    decision-log.md

turns/ is the audit trail. final/report.md is the artifact you share.

Commands

agent-order <scenario-or-file>
agent-order grill <scenario-or-file>
agent-order init
agent-order check
agent-order doctor

Common options:

agent-order ./scenario.md --agents codex,claude
agent-order ./scenario.md --max-turns 10
agent-order ./scenario.md --human-input never
agent-order ./scenario.md --out ./runs

Human Input

Human input is part of the protocol, not a side channel.

Use grill mode when the scenario needs clarification before the Order starts:

agent-order grill "Should we move our frontend to a monorepo?"

This produces human and orchestrator turns:

0001-human.seed.md
0002-codex.intake-question.md
0003-human.intake-answer.md
0004-orchestrator.scenario-freeze.md
0005-codex.initial-position.md

During a run, agents can also emit structured questions for the user. The orchestrator deduplicates them and pauses only when configured.

human_input:
  mode: on_blocking_questions
  max_questions_per_pause: 3

Configuration

Create a starter config:

agent-order init

Default shape:

protocol: agent-order/v1

agents:
  - id: codex
    adapter: codex-cli
    command: codex

  - id: claude
    adapter: claude-cli
    command: claude

limits:
  max_turns: 12

output:
  dir: ./agent-order-runs

synthesis:
  agent: codex

intake:
  enabled: false
  mode: off
  facilitator: codex
  max_questions: 6

human_input:
  mode: on_blocking_questions
  max_questions_per_pause: 3
  ask_before_final: false

final_review:
  enabled: true

If no config sets limits.max_turns, The Order uses max(12, agents * 4 + 4) and reserves room for a synthesis turn.

Adding Another Agent

Codex and Claude are built in. Any scriptable CLI can participate through generic-cli:

agents:
  - id: gemini
    adapter: generic-cli
    command: gemini
    args:
      - -p
      - "{{prompt}}"
    output:
      mode: stdout

The generic adapter can pass prompts through stdin, a prompt file, or templated args like {{prompt}}, then reads either JSON matching the agent-turn schema or plain Markdown from stdout.

Product Positioning

The Order of the Agents is for senior developers, tech leads, staff engineers, and AI-heavy builders who already use terminal AI tools and want better review for consequential decisions:

• architecture choices
• build-vs-buy calls
• migration plans
• security and reliability reviews
• PRD/RFC critique
• incident remediation reviews
• developer-tool product strategy

Use it when the decision is worth a few minutes of critique. Do not use it for every prompt.