How I set up multi-agent workflows in Cursor with reusable skills and agents

Showcase Built for Cursor
,
1

skills subagents multitask

:waving_hand: I built a system that turns Cursor into a full multi-agent orchestration platform. One bash install.sh — 8 specialized agents + 16 workflow skills, ready in every project.

Why I built this

Cursor’s subagent system is powerful but needs setup. Every project, you’re writing agent definitions, figuring out delegation patterns, building workflows from scratch. I wanted /code add auth to just work — planner designs, coder implements, tester tests, reviewer reviews.

What’s included

Command What happens
/code <task> Plan → Implement → Test → Review → Fix
/refactor <task> Analyze → Refactor → Verify → Review
/goal <task> Research → Plan → Execute → Verify until done
/experiment <track> Autonomous experiment loop (Karpathy’s autoresearch)
/pr, /review, /commit, /test Daily workflow automation
/comps, /dcf, /lbo Financial modeling → Excel output
/init Scan project → Generate context rules
/doctor Verify installation health + skill routing

Architecture — two building blocks

The whole system is built on just two primitives that compose freely:

  • Agents (~/.cursor/agents/*.md) — Role definitions: planner, coder, reviewer, tester, researcher, architect, financial-analyst, codebase-explorer
  • Skills (~/.cursor/skills/*/SKILL.md) — Workflow definitions that orchestrate agents through phases

Skills define what to do — Agents define who does it.

/code skill workflow:

  • Phase 1: Plan → spawns planner agent
  • Phase 2: Implement → spawns coder agent
  • Phase 3: Test → spawns tester agent
  • Phase 4: Review → spawns reviewer agent

Mix and match freely: use built-in agents in custom skills, or custom agents in built-in skills.

Each workflow includes:

  • Phase verification — supervisor confirms each phase produced valid output
  • Execution tracking — structured summary printed at the end
  • Fault recovery — defined strategies for each failure mode (e.g., coder breaks tests → revert + retry with error context)

Extending it

Adding a new agent: one .md file with a role prompt (30 seconds). Adding a new workflow: one SKILL.md with phase definitions (2 minutes). I’ve built up to 8 agents and 16 workflows covering coding, refactoring, testing, and financial modeling.

I packaged the whole thing so it installs with one command. Repo is here if anyone wants to try it or build on it: GitHub - cocolwy/cursor-agents: MCCA (Make your Cursor Claude Again) · GitHub

Would love to hear how others are handling multi-agent setups — especially phase transitions and error recovery.