Skip to content
Agent Squad

Building with an AI assistant

The repo ships SKILL.md — a single, assistant-agnostic guide at the root of the swift/ directory. It covers:

  • Mental model — when to reach for Agent vs. GroundedAgent, single-agent vs. classifier routing, text vs. voice
  • Real API signatures — Orchestrator, Agent, GroundedAgent, LLMClassifier, ChatCompletionsClient, storage and tracing types
  • Module map — what each import (AgentSquad, AgentSquadMCP, AgentSquadAudio) pulls in
  • Task recipes — the AsyncThrowingStream<AgentEvent, any Error> loop, routing between agents, wiring tools, custom protocol conformances
  • Gotchas — maxToolRounds defaults, JSONValue integer decoding, storage scoping, tracing lifecycle, realtime teardown

Pointing your assistant at SKILL.md

Section titled “Pointing your assistant at SKILL.md”

Tell your assistant once, before it writes any AgentSquad code:

Read SKILL.md before writing any AgentSquad code.

That is all. The file is self-contained — no context beyond it is required.

This works with any assistant: Claude, Cursor, GitHub Copilot, or anything else that can read a file.

Claude Code: install as a skill

Section titled “Claude Code: install as a skill”

Claude Code users can register SKILL.md so it is loaded automatically whenever the assistant works on AgentSquad code. Copy the file into your project’s skills directory:

Terminal window
mkdir -p .claude/skills/agent-squad-swift
cp swift/SKILL.md .claude/skills/agent-squad-swift/SKILL.md

After that, Claude Code picks it up without an explicit instruction on every session.

What SKILL.md does not replace

Section titled “What SKILL.md does not replace”

SKILL.md is guidance and a map, not an API reference. It tells you what to use, when, and what to watch out for. For exact signatures, read the sources under swift/Sources/AgentSquad/. For worked examples, see the rest of this doc site — in particular: