Writing a collaboration skill

Tools let agents talk; a skill teaches them how to collaborate well.

Why a skill

The Agent Bus gives agents the abilityto message each other. A short skill — a markdown file dropped into your agent's instructions (e.g. CLAUDE.md, AGENTS.md, or a Claude Code skill) — teaches the protocol: read before you write, address by session_id, one message then wait, and don't loop forever.

The protocol

  • Listlist_agents to find the target by role and grab its session_id.
  • Read firstread_pane the target so you don't talk over it.
  • Message — one clear, self-contained request; identify yourself and say what you want back.
  • Read again — wait, then read_pane for the reply. Re-read rather than spamming.

Etiquette & safety

  • One message, then wait — multiple messages before a reply garble the other agent's input.
  • Be explicit about scope — other agents may be in autonomous/YOLO mode and will act on what you send. Say "propose only" when you just want analysis.
  • Don't loop — converge on a result and report to the user instead of infinite ping-pong.
  • Read-only when unsureread_pane is always safe for observing a long-running agent.

Drop-in skill

CodeGrid ships a ready-made one at docs/agent-bus-skill.md in the repo. Add it to your agent's context, or paste this frontmatter-style header to make it a Claude Code skill:

agent-bus-skill.md (header)
---
name: codegrid-agent-bus
description: How to collaborate with other agents in CodeGrid via the
  codegrid-agent-bus tools (list_agents, read_pane, message_agent).
  Use when asked to delegate to, consult, or coordinate with another agent.
---
TIP
Keep the skill short and imperative. Agents follow a tight "list → read → message → read" loop far more reliably than a wall of prose.