Dynamic multi-agent workflow consumes full credit balance while ~60% of agents never run

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Two related problems, both around extra usage credits and multi-agent
workflows.

1. Purchased credits do not lift the weekly-limit block, but are still

consumed.

After hitting the weekly usage limit, I purchased USD 200 of prepaid extra
usage credits because the CLI itself told me to ("/usage-credits to finish
what you're working on"). After the purchase, the message
"You've hit your weekly limit · resets Sep 23, 1 pm (Asia/Tokyo)" continued to
appear at every subsequent step, while the balance was drawn down to zero.
The credits did not restore the access they were sold to restore.

1. A dynamic multi-agent workflow consumed the entire balance while most of

its agents never ran.

Claude launched a workflow that fans out many parallel verification agents.
It ran for about 1 hour 50 minutes and was cut off; a second attempt ran ~22
minutes; a third terminated after 2 seconds. Claude reported that roughly
60% of the verification agents never executed. Agents that are launched and
then terminated appear to be billed in full, and the partial results that did
come back were largely discarded.

There is no cost estimate before fan-out, no cap on how much a single
workflow invocation can consume, and no warning that one workflow could
exhaust a USD 200 balance in under two hours. The net result of USD 200 was
an incomplete verification pass.

Anthropic Support has declined a refund (Invoice 8QKUCYZ8-0017, conversation
ID 215476005043016), so I am filing this as a product issue rather than a
billing one.

What Should Happen?

Extra usage credits, purchased in response to the CLI's own prompt to
"finish what you're working on", should actually restore access when the
weekly limit has been hit. If the weekly limit is still going to be
enforced, the credits should not be drawn down at all.

A dynamic multi-agent workflow should:

• show an estimated cost and require confirmation before fanning out a large

number of parallel agents;

• stop cleanly when the remaining balance cannot support the planned fan-out,

instead of launching agents that cannot complete and billing for them;

• preserve and return the results of agents that did finish, rather than

discarding them when the run aborts.

In my case, USD 200 was consumed in under two hours and produced an
incomplete verification pass in which roughly 60% of the agents never ran.
None of the above safeguards were present.

Error Messages/Logs

You've hit your weekly limit · resets Sep 23, 1 pm (Asia/Tokyo)
/usage-credits to finish what you're working on.

(Displayed repeatedly after extra usage credits had already been purchased:)

✻ Brewed for 1h 50m 28s · done 5:33
● Dynamic workflow "Re-verify the 9/19 proposal after fixes: robotics story,
  call alignment, layout, facts, advice coverage vs transcript, regression"
  completed · 20m 35s
You've hit your weekly limit · resets Sep 23, 1 pm (Asia/Tokyo)

✻ Churned for 2s · done 13:48
You've hit your weekly limit · resets Sep 23, 1 pm (Asia/Tokyo)

Steps to Reproduce

1. Use Claude Code on an Individual plan and continue working until the

weekly usage limit is reached. The CLI displays:
You've hit your weekly limit · resets <date> · /usage-credits to finish
what you're working on.

1. Purchase prepaid extra usage credits (USD 200 in my case) as instructed

by that message.

1. Tell Claude that credits are now available and ask it to resume the task

in progress. In my case, the task was verifying a proposal document across
several dimensions (story consistency, alignment with meeting notes,
layout, factual consistency).

1. Claude launches a dynamic multi-agent workflow that fans out many parallel

verification agents.

Observed:

• The workflow ran for approximately 1 hour 50 minutes and was cut off.

The weekly-limit message was displayed throughout, even though credits had
just been purchased.

• A second attempt ran for about 22 minutes and hit the same limit message.
• A third attempt terminated after 2 seconds with the same message.
• Claude reported that approximately 60% of the verification agents never

ran, and fell back to manually inspecting the few surviving results.

• The entire USD 200 balance was consumed across these runs.

Expected:

• Extra usage credits should restore access when the weekly limit is hit,

or should not be drawn down while that limit is still being enforced.

• A workflow should not be able to consume an entire credit balance in under

two hours with no cost estimate before fan-out and no budget guardrail.

• Results from agents that did complete should be preserved and returned

rather than discarded when the run aborts.

Note: This is not reproducible with a minimal code example, because the
trigger is the interaction between the weekly-limit state, a credit purchase,
and a multi-agent workflow fan-out — not a specific file or language.
Anthropic Support has the session on record (Invoice 8QKUCYZ8-0017,
conversation ID 215476005043016).

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Code 2.1.278

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_