Multi Agents

An exploration of multi-agent workflows for development environments.

In plain language: this is a setup that lets you talk to your coding assistant normally, while a small agent system decides whether the task needs a quick answer, a plan, or delegated work. The goal is to make AI-assisted development feel simpler for the human while keeping model cost, context size, and implementation risk under control.

Architecture

The repo is organized in three layers:

multi-agents/
├── agents/          # Tool-agnostic agent definitions (model placeholders)
├── skills/          # Reusable knowledge modules loaded on demand
├── opencode/        # OpenCode-specific config (AGENTS.md, commands, opencode.json)
├── claude/          # Claude Code-specific config (CLAUDE.md, commands)
└── distribute.sh    # Sync agents, skills, and commands to OpenCode / Claude Code

Agents

Six hidden workers + two visible orchestration agents:

Agent	Model	Role
worker-lite	Haiku	Cheap read-only work — file lookup, log reading, summaries
worker-standard	Sonnet	Normal implementation, debugging, tests, refactors
worker-heavy	Opus	Architecture, security, subtle bugs, high-risk decisions
planner	Sonnet	Creates plans for user approval before execution
orchestrator	Sonnet	Executes approved plans by delegating to the cheapest capable worker
agent-manager	Haiku	Creates and updates agents, skills, and global config

Skills

Domain knowledge loaded on demand — no model or permissions, pure content:

Skill	Use for
git-workflow	Branch naming, conventional commits, PR rules, gh CLI
code-review	Per-change analysis, review checklist, LGTM/Block verdict
documentation	Structured Markdown notes saved to Obsidian vault
claude-api	Claude API patterns, prompt caching, tool use, cost optimization
agent-creator	Design and write new agent files
skill-creator	Design and write new skill files

Request routing

User request
    ↓
Trivial (1 step)      → Primary agent answers directly
Moderate (2-3 steps)  → worker-lite (reads) + worker-standard (writes)
Complex               → @planner → user approval → @orchestrator → workers

Install

# First time (OpenCode config + AGENTS.md)
./opencode/scripts/install.sh

# Sync agents, skills, and commands (run after any change)
./distribute.sh

# OpenCode only or Claude Code only
./distribute.sh --opencode-only
./distribute.sh --claude-only

Model overrides

The default models use GitHub Copilot. Override via env vars:

# Example: use OpenAI models instead
MODEL_LITE=openai/gpt-4o-mini MODEL_STANDARD=openai/gpt-4o MODEL_HEAVY=openai/o3 ./distribute.sh

Default model IDs:

Placeholder	Default
{{MODEL_LITE}}	github-copilot/claude-haiku-4.5
{{MODEL_STANDARD}}	github-copilot/claude-sonnet-4.6
{{MODEL_HEAVY}}	github-copilot/claude-opus-4.6

Claude Code

Copy or symlink claude/CLAUDE.md into your project:

cp claude/CLAUDE.md /path/to/your-project/CLAUDE.md

The slash commands in claude/commands/ are synced to ~/.claude/commands/ by distribute.sh.

Cost Metrics

The repository defines a cost-aware model-routing policy and tracks routing efficiency.

Pilot result, 2026-04-26:

Worker	Calls	Input	Output	Heavy model avoided
worker-lite	3	3.4K	111	Yes
worker-standard	4	12.3K	575	Yes
worker-heavy	0	0	0	—

77% cost reduction vs. all-Sonnet baseline on multi-step tasks. Detailed report: Model Routing Efficiency Report.

Environments

• opencode/ — OpenCode environment with OpenAI models (first pilot)
• agents/ + skills/ — Tool-agnostic layer, uses Claude models via distribute.sh