### Preflight Checklist - [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Amodel) for similar behavior reports - [x] This report does NOT contain sensitive information (API keys, passwords, etc.) ### Type of Behavior Issue Other unexpected behavior ### What You Asked Claude to Do Run a small refactoring task on using Fable 5 with Ultracode mode enabled. ### What Claude Actually Did **What Claude Actually Did** 1.

[MODEL] Fable 5 Ultracode spawns excessive parallel agents for single refactoring task

Status Fixed / completed

Maintainer reply None cached

Workaround ✓ Mentioned in thread ↓

Activity 4 comments · opened Jun 10, 2026 · closed Aug 25, 2026

Preflight Checklist

[x] I have searched existing issues for similar behavior reports
[x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

Run a small refactoring task on using Fable 5 with Ultracode mode enabled.

What Claude Actually Did

What Claude Actually Did

Fable 5 spawned 7 parallel agents for a single refactoring task
Agent 1 consumed 139.9k tokens
Agent 2 consumed 108.7k tokens
Agents 3-5 consumed 77-91k tokens each
Agent 6 consumed 90.4k tokens
Agent 7 consumed 128.6k tokens
Total consumption: 716.6k tokens in 5m 35s
Session quota jumped from 33% → 82% in single workflow, then to 90% total

Expected Behavior

Fable 5 should spawn agents conservatively and proportionally to task complexity, similar to Opus 4.8 and other models. A single refactoring task should not require 7 parallel agents. Quota consumption should be proportional to actual work performed.

Files Affected

No unexpected file modifications observed

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Open Claude Code app on Windows
Set model to Fable 5 with Ultracode mode enabled
Request a refactoring task (e.g., "Refactor map-composer-subsystem")
Observe agent spawning behavior in real-time
Check /usage to see token consumption breakdown

Claude Model

Other

Relevant Conversation

Fable 5 spawned agents without explicit instruction to do so. The model appears to be over-interpreting task complexity and creating unnecessary parallel workers.

Impact

Low - Minor inconvenience

Claude Code Version

1.11847.5 (9692f0) 2026-06-09T17:34:27.000Z

Platform

Anthropic API

Additional Context

High Quota consumption is disproportionate to work performed. Same task on Opus 4.8 consumed ~1M tokens but only used 33% of quota. Fable 5 consumed 716.6k tokens and used 49% of quota in a single workflow.

Pattern observed:

Fable 5 with Ultracode mode consistently spawns excessive parallel agents
Opus 4.8 with Ultracode mode does NOT spawn 7 agents on the same task type
This behavior is specific to Fable 5 and appears to be a model-level decision, not a configuration issue

View original on GitHub ↗

3 Comments

DrAlexHarrison · 2 months ago

Agreed. The agent-spawn recklessness bug is legit in fable. Highly recommend blocking that in CLAUDE.md. It listens to my hard cap of 10 agents at a time, and friendly reminder that each agent is mostly functional on it's own out to 500-800k tokens and doing things serially is more token-efficient.

Also funny that Fable literally will (unregulated in CLAUDE.md) spawn so many agents as to get auto-rate-limited by anthropic and burn a few million tokens for nothing when they crash due to throttling.

kcarriedo · 2 months ago

This is exactly the right signal to file. The uncontrolled agent proliferation in Fable 5 / Ultracode mode is a real coordination failure — 7 agents spawned for a single small refactoring task means there's no proportionality signal flowing from the orchestrator to the model about what scale of parallelism the task actually warrants.

DrAlexHarrison's workaround (hard cap in CLAUDE.md) is pragmatic but it's a manual governor on top of a missing architectural constraint. The model has no feedback loop from the actual work scope back to its spawning decision — it fires up agents speculatively, burns quota, and the user only discovers the overshoot at /usage.

A few things that help in the interim while this gets fixed at the model level:

Explicit task-scope framing in your prompt — "This is a single-file refactor, do not spawn parallel agents" before triggering Ultracode. Verbose, but Fable 5 does respect explicit scoping instructions in our testing.

CLAUDE.md hard cap (as DrAlexHarrison notes) — max_parallel_agents: 3 or similar.

Pre-flight cost estimate — Opus 4.8 without Ultracode for scope estimation, then decide whether to escalate to Fable 5 Ultracode. The 2-step cost is less than a 716k token surprise.

The deeper fix is proportional spawn logic — the orchestrator should reason about task complexity (file count, LOC, cross-module surface area) before deciding agent fan-out, rather than defaulting to maximum parallelism. Looking forward to seeing Anthropic address this; the token-proportionality expectation in the OP is exactly right.

cantonalex · 1 month ago

How this is a shipped project is kinda insane

Showing cached comments. Read the full discussion on GitHub ↗