#claude-agents

Multi-agent patterns for Claude Code: a strong model (Fable 5) supplies judgment, a fast model (Sonnet) does the mechanical work.

#Install

Copy the .md files into .claude/agents/ in your project (or ~/.claude/agents/ for all projects).

sonnet-executor and fable-orchestrator are entry points — run them as your main agent (they need the Task tool to spawn the others). fable-advisor and sonnet-worker are the agents they spawn; you don't normally invoke those directly.

#Design notes

The prompting in these agents adapts the measured guidance from Anthropic's official advisor tool documentation to Claude Code subagents:

• Call timing: consult before substantive work (orientation reads don't count), once before committing to an approach and once before declaring done. Make the deliverable durable before the done-check.
• Advice weight: follow the advice unless it fails empirically; on a conflict between gathered evidence and the advice, do one reconcile consult instead of silently switching.
• Advisor brevity: advisor output length is the pattern's main cost driver, so the advisor is instructed to stay under ~200 words.

One structural difference from the API advisor tool is load-bearing: the API advisor automatically receives the executor's entire transcript; a Claude Code subagent sees only the prompt it's handed. That's why sonnet-executor is required to pass task + intent + evidence + file paths on every consult, and why fable-advisor reads the named files itself before advising.

Other Claude Code realities baked in:

• Subagents share nothing — no history, no prompt cache, no knowledge of sibling sub-tasks. The orchestrator writes self-contained worker prompts and owns cross-file consistency.
• Parallel workers must not edit the same file; same-file sub-tasks run sequentially.

#Building on the API instead?

If you're building your own agent on the Messages API rather than Claude Code, use the server-side advisor tool — the executor pauses mid-generation, the advisor sees the full transcript automatically, and everything happens in one request:

response = client.beta.messages.create(
    model="claude-sonnet-4-6",              # executor
    max_tokens=4096,
    betas=["advisor-tool-2026-03-01"],
    tools=[{
        "type": "advisor_20260301",
        "name": "advisor",
        "model": "claude-opus-4-8",          # advisor — must be ≥ executor capability
        "max_tokens": 2048,                  # caps advisor output (recommended starting point)
    }],
    messages=[{"role": "user", "content": "..."}],
)

Notes from the official docs: the advisor model must be at least as capable as the executor (invalid pairs 400); a Fable 5 executor can only be advised by Fable 5, and Fable/Mythos advisors return advisor_redacted_result (encrypted — round-trip it verbatim); pass advisor_tool_result blocks back on subsequent turns, and if you later drop the advisor tool, strip those blocks from history too or the request 400s; advisor usage is billed at the advisor model's rates and reported per-iteration in usage.iterations[]; enable the tool-level caching option only when you expect ≥3 advisor calls per conversation.