Agent Lab

Agent Lab

CLI tool to orchestrate teams of Claude Code agents. Define agents with different roles in a YAML file, Agent Lab runs them in parallel respecting dependencies, manages git branches, merges, and collects results.

Installation

# Base (CLI only)
pip install -e .

# With Web UI
pip install -e ".[web]"

# With internet access for agents (search, fetch, scrape, API calls)
pip install -e ".[internet]"

# Everything
pip install -e ".[web,internet,dev]"

Requires [Claude Code](https://docs.anthropic.com/en/docs/claude-code) installed and configured (`claude` available in PATH).

For internet-enabled research, optionally run [SearXNG](https://github.com/searxng/searxng) in Docker for web search:

docker run -d --name searxng -p 8888:8080 searxng/searxng

Quick Start

# Create a sample team
agent-lab init

# Run
agent-lab run team.yaml --verbose

# Or launch a single agent on the fly
agent-lab solo "security expert" "Review ./src for vulnerabilities"

# Web UI
agent-lab web
# Open http://127.0.0.1:8420

---

Key Concepts

Agents

Each agent has a **name**, a **role**, and a **mode**:

Mode	Behavior	Default tools
code	Reads and modifies files, does work	Read, Write, Edit, Glob, Grep, Bash
report	Analysis only, does not modify files	Read, Glob, Grep, Bash

Teams

A team groups agents with a common **task**. Three strategies:

• pipeline (default) — Agents run in waves. Dependencies (depends_on) determine order. Agents in the same wave run in parallel.
• orchestrated — An orchestrator agent dynamically decides which workers to call and in what order.
• adaptive — A goal-driven loop with controller, evaluator, and shared state.

Waves and Dependencies

Wave 1: [planner]