Skip to content

2026-04-24-multi-agent-direction.md

Latest commit

 

History

History
78 lines (55 loc) · 4.86 KB

2026-04-24-multi-agent-direction.md

File metadata and controls

78 lines (55 loc) · 4.86 KB
Raw

Direction: Cockpit is Multi-Agent

Date: 2026-04-24 Status: Direction statement — guides all future work Related issues: #31 projection, #32 Karpathy skill, #33 Multica reference

Decision

Cockpit is a multi-agent orchestration layer, not a Claude Code accessory. It should support Claude Code, Codex, Cursor, Gemini CLI, and future coding agents equally. Features that only work for Claude Code are a migration target, not a permanent state.

Why now

The plugin system (phases 1-4, all shipped) already abstracts the four major Claude-specific surfaces behind driver interfaces:

  • Runtime — cmux today; tmux / Docker / SSH / other CLIs are future drivers
  • Workspace — Obsidian today; Notion / plain-md / S3 are future drivers
  • Tracker — GitHub today; Linear / Jira / GitLab are future drivers
  • Notifier — cmux today; Slack / Discord / email / pager are future drivers

The abstraction proved out. The missing layer is agent-side: templates, skills, and instructions that only Claude Code reads.

What "supports all" means concretely

Layer Claude Code Codex Cursor Gemini CLI
Runtime ✅ cmux ✅ (runtime driver) ✅ (runtime driver) ✅ (runtime driver)
Instructions CLAUDE.md AGENTS.md .cursor/rules/*.mdc GEMINI.md
Skills SKILL.md via Skill tool via AGENTS.md include via rule include via GEMINI.md include
MCP config ~/.claude/settings.json ~/.codex/config.toml ~/.cursor/mcp.json varies
Memory claude-mem (MCP) claude-mem (MCP, if MCP supported) claude-mem (MCP) claude-mem (MCP)

Pain points when using non-Claude agents today

From trying Cursor / Codex / Gemini CLI on cockpit-managed projects:

  • Skills written for Claude's Skill tool aren't invoked (other agents don't have the tool)
  • CLAUDE.md rules aren't read (agent looks at its own file)
  • MCP servers aren't configured (each agent has its own config path)
  • Daily logs / standup / retro integrations assume cmux + Claude
  • Captain/crew templates are written Claude-specific (.claude.md) — generic versions (.generic.md) exist but are thinner

Strategy

Principles

  1. Claude Code is still the reference implementation. It ships working today; other agents are migrations, not rewrites.
  2. AGENTS.md is the canonical instruction format. CLAUDE.md becomes a thin wrapper or symlink.
  3. Skills stay portable markdown. Claude Code reads via the Skill tool; other agents read via AGENTS.md inclusion. Don't translate frontmatter until a cross-agent SKILL.md spec stabilizes.
  4. Delegate what others already own. MCP config sync → conductor or mcp-linker. Skill sharing → skillshare. Don't reinvent.
  5. Don't copy Multica. Their team + cloud model is a different product. Cockpit's moat is solo-dev + knowledge-graph + local-first.

Sequence

  1. Now — direction stated (this doc, README, CLAUDE.md)
  2. Next — Karpathy skill ported + wired into captain/crew templates (#32)
  3. Soon — cockpit projection command for cross-agent config sync (#31)
  4. Later — evaluate Multica ideas (#33), runtime drivers for non-cmux agents

Non-goals

  • Team collaboration features (that's Multica's space)
  • Web UI / daemon / WebSocket streaming (local-first, terminal-first)
  • Replacing cmux as the default runtime (cmux stays the reference)
  • Waiting for cross-agent SKILL.md standardization before shipping anything

Success criteria

  • A dev can open a cockpit-registered project in Codex, Cursor, or Gemini CLI and get roughly equivalent context (role identity, skills, MCP servers).
  • "Roughly equivalent" means: the agent knows it's a captain/crew, knows how to spawn subtasks, has access to the same MCP servers, and can read status/learnings/wiki.
  • Claude Code experience does not regress.

Inspirations