Multi-Agent Setup

Configure multiple agents with delegation, scoped tools, and specialized system prompts.

A single agent handles most tasks. When you need specialization — a researcher that delegates writing, a coordinator that dispatches to experts — set up multiple agents with delegation.

1. Define agents in crab.toml

[system.crab]
model = "deepseek-chat"

[agents.researcher]
description = "Deep research specialist"
model = "claude-sonnet-4-20250514"
max_iterations = 32
thinking = true
members = ["writer"]
skills = ["web-research"]
mcps = ["search"]

[agents.writer]
description = "Writes clear, concise content"
model = "deepseek-chat"
skills = ["technical-writing"]

The members field lets the researcher delegate tasks to the writer. Each agent can have its own model, iteration limit, and scoped access to skills and MCPs.

2. Write system prompts

Create ~/.crabtalk/agents/researcher.md:

You are a research specialist. When given a topic:
1. Search for authoritative sources
2. Synthesize findings into structured notes
3. Delegate writing tasks to the writer agent

Create ~/.crabtalk/agents/writer.md:

You are a technical writer. Take research notes and produce
clear, well-structured content. Be concise.

3. Test delegation

crabtalk attach researcher

> Research the current state of WebAssembly and write a summary

The researcher searches, gathers information, then spawns a task for the writer to produce the final output.

Scoping

Each agent only sees the skills and MCPs listed in its config. Empty lists mean access to everything. Use scoping to keep agents focused:

[agents.researcher]
skills = ["web-research"]     # only this skill
mcps = ["search"]             # only search tools

[agents.writer]
skills = ["technical-writing"]
mcps = []                     # no MCP tools

What's next

Agents — full agent configuration reference
Skills — how skills work and how to scope them