Code8545

Fable as the manager: Claude Code subagents

Link

Guide

Make Fable the manager: Claude Code subagents

Start your Claude Code session with Claude Fable 5.2 and tell it "you are the manager". Fable plans the work, gives subtasks to Opus or Haiku subagents and checks the results. Each subagent works in its own context and sends back only a summary, so Fable, the expensive model, keeps a small context while the real work runs on cheaper models.

1. The manager prompt (copy it)

Open Claude Code with Fable (pick it with /model) and write this at the start of the project:

You are the manager of this project. Run Opus or Haiku subagents in sequence for coding, analysis and other subtasks: give hard, multi-step work (architecture decisions, complex code) to Opus, and simple, repetitive work (searching files, reading, summarizing, running tests) to Haiku. You only split the work, give each agent a clear task and check what they did. Don't write the code yourself; be the manager over them.

2. Subagent files (the safer way)

The prompt works on its own, but fixing which agent runs on which model in a file is more reliable. Put two files in .claude/agents/ in your project (or ~/.claude/agents/ for all your projects):

.claude/agents/worker.md

---
name: worker
description: Writes and edits code for a clearly defined task from the main session. Use for implementation work.
model: opus
---

You implement the task you are given. Read the files you need, make the change, run the relevant checks and report back with a short summary of what you changed and anything that failed.

.claude/agents/scout.md

---
name: scout
description: Finds files, reads code and docs, runs tests and summarizes the results. Use for search, reading and simple checks.
tools: Read, Grep, Glob, Bash
model: haiku
---

You look things up and report back. Keep your answer short: the file paths, the lines that matter and a one-paragraph summary.

Then name the agents in your manager prompt: "Use scout for research and tests, and worker for code."

3. Which job goes to which model

  • Fable 5.2 (main session): planning, splitting the work, checking.
  • Opus 5.5: hard code, architecture decisions, multi-step reasoning. The docs: "Reserve Opus for complex architectural decisions or multi-step reasoning."
  • Haiku 5.5: simple work like searching, reading files, summaries and running tests. The docs: "For simple subagent tasks, specify model: haiku in your subagent configuration."

4. Prices (per million tokens, November 20, 2026)

ModelInputOutput
Claude Fable 5.2$8$40
Claude Opus 5.5$4$20
Claude Haiku 5.5 (prompts up to 100,000 tokens)$0.10$0.50

Good to know

  • Subagent requests count toward the same usage limits. You save because the work runs on cheaper models and Fable's context stays small; the total work does not shrink.
  • A subagent's model is set in this order: the model Claude passes when it starts the subagent, the model field in the agent file, the CLAUDE_CODE_SUBAGENT_MODEL environment variable, and last the main session's model. With no model in the file and none passed by Claude, the subagent runs on Fable, so writing model: opus / model: haiku is the safest way.
  • When the main session runs Fable, the built-in Explore agent runs on Opus with a Claude subscription or a Console account.
  • Many subagents returning long results can still fill the main context; ask them for short summaries.

Source: Claude Code docs, code.claude.com/docs/en/sub-agents and code.claude.com/docs/en/costs; prices from platform.claude.com/docs/en/about-claude/pricing.

Looking for a different code?

Sponsored

YoodaAI logoYoodaAIAdLive AI answers in your interview, plus mock practice.

This spot is open to sponsors

Show your brand to 5K+ visitors a month.

hi@codebasehub.pro