ASO Dual-Agent Review Kit

By ASO Ltd. — the digital backbone of African business.

Make two AI agents review each other's work before you ship it — without you playing messenger between them.

Works for any kind of work, not just code: documents, marketing content, contracts, financial models, designs, websites. Any two agents: Claude Code + OpenAI Codex, Claude + Gemini CLI, two Claude projects — if it can read a folder and follow an operating manual, it can run this loop.

This is the exact system ASO Ltd. uses in production: every change to our own website is proposed by one agent, adversarially reviewed by the other, and merged only by a human after both sign off. It has caught real bugs neither the author-agent nor a single review would have — dead fallback paths, self-contradicting page copy, a spam honeypot that looked right but could never fire.

One agent reviewing its own work has a blind spot: the same reasoning that produced the mistake tends to miss it. A second, independent agent — different model, fresh context — catches what the first cannot.

But naïve two-agent setups turn you into the bottleneck: you relay "the other one reviewed," re-explain the protocol every session, and babysit every round-trip.

How the loop works (and why it's plain files, not n8n or Zapier):

This kit fixes both:

1. A paper-trail protocol — proposals, reviews, revisions, and sign-offs live as markdown files in a shared folder. Every decision is auditable.
2. Silent whispers — each agent runs a tiny script at session start that prints one line only when a proposal is waiting on that agent, and nothing otherwise. You open an agent; if it's that agent's turn, it already knows. Near-zero token cost.
3. Hard rules baked into every file — agents never ship/merge/publish (you do), never approve their own work, never rubber-stamp. The turn-tracking logic guarantees a proposal is never "waiting on" both agents at once.

1. Copy .claude/skills/dual-agent-review/ from this repo into your project's .claude/skills/ folder.
2. Open the project in Claude Code and say: "set up dual-agent review".
3. Answer the ~6-question interview: which two agents, what kind of work, where work-in-progress lives (git branches or draft files), what the human's final act is (merge / publish / send / sign), and what the reviewer should check against.
4. The skill installs everything, adapted to your answers, and verifies the whispers before declaring done.

Day-to-day after that:

• You: "make X change" → Agent A does it + writes a proposal, stops.
• You open Agent B → its whisper says a proposal awaits → it reviews for real.
• Changes requested? Agent A's next session whisper tells it. Fixed, re-reviewed, both agree.
• You perform the final act. That's the only part that's yours — as it should be.

On the Claude Code side you also get /check-reviews — one phrase that scans the queue, classifies whose turn it is, and does whatever this agent owes: fix, review, or countersign.

No — nothing inside Codex itself. No config.toml edit, no plugin, no toggle. Setup runs once from Claude Code (that's where the skill lives), and it provisions both sides by writing plain files. For Codex it writes two:

1. Codex's operating manual (AGENTS.md) — the review-loop rules, plus a first line telling Codex to run its whisper at session start.
2. Codex's whisper script — the bash file that checks the folder.

Codex reads AGENTS.md automatically — that's its own native convention — so there is nothing to wire up on the Codex side.

The only prerequisite isn't kit setup, it's just having the second agent: Codex CLI installed and logged into your own account. Same for any other second agent (Gemini CLI, a second Claude project).

One honest asymmetry. On the Claude Code side the whisper is a hook — it fires automatically every session, guaranteed by the harness. Codex has no hook system, so it runs its whisper because its manual tells it to, the moment you open Codex and send a message. Very reliable in practice; if it ever skips, "run your session-start check" nudges it back. Two agents is the sweet spot precisely because at least one of them (Claude) enforces the check automatically.

• The proposal file format is the stable interface. Proposed by:, ## Review — <name>, ## Revision — <name>, and two AGREED checkboxes. Anything that reads/writes these stays compatible — customize everything else freely.
• The proposer must declare what it verified and where it's least confident. That one field is worth more than the rest of the proposal combined.
• Whispers are exhaustively turn-aware. Proposal awaiting first review → only the reviewer hears. Changes requested and unaddressed → only the proposer. Revised → only the reviewer. Approved awaiting countersign → only the proposer. Done → silence. Tested on macOS's default bash 3.2.
• Two agents, not three. A third reviewer triples cost for diminishing returns. The human is the third reviewer.

Installing the kit takes minutes. Designing the workflows around it — what your agents should build, review, and automate across payments, logistics, and operations — is what ASO Ltd. does for clients across Cameroon and Central Africa. AI & Intelligent Automation is one of our eight service lines, built for XAF, MTN/Orange Mobile Money, French and English.

• Talk to us: consult@asoltd.info — we respond within 24 hours
• See what we build: www.asoltd.info

More free tools from ASO: AIOS Starter — Claude Edition · AIOS Starter — Codex Edition

MIT License. © 2026 ASO Ltd.