pair-agent-harnesslisted

Harness two AI agents (typically Claude + OpenAI Codex, or two Claudes, or an agent + human reviewer) into a Reviewer / Implementor pair that collaborate asynchronously through a shared markdown thread. One agent reviews / critiques / asks pointed questions; the other implements / responds / drives the code changes; the file mediates. Use when the user says "pair another agent on this", "harness Codex on this", "review this with Codex", "open a Codex thread", "respond to Codex", "open a peer review", "monitor this review file", names a *-review-thread.md or *-dialogue-thread.md, or wants to coordinate structured async back-and-forth between two AI agents on a code or design question. Seeds the file from a template if missing, appends a numbered Claude response under its own section header, arms a file-mtime monitor so each peer append triggers a fresh read + response, drives the pyramid-principle response shape (position first, justification second), detects consensus markers like "Final ACK" / "No further ob
Rijul1204/rashedul-agentic-engineering · ★ 1 · AI & Automation · score 72

Install: claude install-skill Rijul1204/rashedul-agentic-engineering

# Pair-Agent Harness A coordination harness for two AI agents working in **complementary roles** on a shared markdown file. The two canonical roles are: - **Reviewer** — critiques, challenges assumptions, asks pointed questions, ratifies decisions. - **Implementor** — proposes designs, implements code changes, responds to critique, runs quality gates. The roles swap per topic — when Claude opens the thread with findings + questions for Codex, Claude is the Implementor / Codex is the Reviewer. When Codex opens the thread asking for review of a Codex-generated patch, the roles flip. The harness doesn't pin roles; it pins the **discipline**: one agent appends at a time, each side reads the whole file before responding, the markdown thread is the single source of truth. This decouples the two agents from each other's runtime: neither has to be online at the same time. The Implementor can land code changes between rounds; the Reviewer can take a day to think; the file persists across both sessions. ## When to invoke Fire on: - User says "pair another agent on this", "harness Codex on this", "pair me with Codex on this design". - User says "review this with Codex", "open a Codex review", "let's get Codex's take on this". - User says "open a peer review", "get a second pair of eyes on this", "ping another Claude on this". - User points at an existing `*-review-thread.md` / `*-dialogue-thread.md` and says "respond" / "continue" / "monitor". - User says "monitor this file" with