Modern software development is fractal. Every feature touches 5-10 contexts
simultaneously. A single LLM struggles to maintain multi-domain coherence.
This is the problem Agent Swarm solves.
9 Agents, One Pipeline
Unlike single-agent tools, Swarm is a pipeline where 9 specialized AI agents
collaborate automatically, each with a precise role and strict constraints:
Orchestrator, Search, Planner, Contract, Front, Back, Tester, Reviewer, Writer.
Automatic Classification into 5 Routes
The Orchestrator uses mandatory pre-search: extracts technical terms, runs
parallel grep on the codebase, and counts distinct files touched. No LLM
guessing its own difficulty. Routes: DIRECT → SIMPLE → ADAPT → MEDIUM → FULL.
Real Example: 8 minutes vs 2-4 hours
"Add a Recent Updates section to the homepage" → 8 files touched → MEDIUM.
Planner designs 5-step plan. Front+Back implement in parallel. Tester ensures
80% coverage. Reviewer audits (security >=1.0, quality >=0.85). Merge automatic.
Results
- 5-8x cheaper than single-agent (specialization reduces context 80%)
- MIT licensed, built for OpenCode
- Full wiki: https://swarm-wiki.vercel.app
3-Minute Setup
git clone https://github.com/JohTandou/agent-swarm.git
cp -r agent-swarm/.opencode/* ~/.opencode/
cp agent-swarm/opencode.json ~/.config/opencode/opencode.json