Agent-initiated multi-agent runs exhaust the 5-hour window with no estimate and no warning

Plan: Max · Claude Code in the desktop app.

What happened

During a delivery session, Claude Code launched a multi-agent review on its own initiative — six reader agents, their verifier agents, and 34 mutation runs on a copy of the repo. That single decision exhausted my 5-hour window. The review was killed before writing its synthesis; its 74 sub-results survived only because they had been written to the workflow journal and could be recovered by hand.

At that moment: 5-hour window 100%, weekly allowance 32%. The work stopped with two thirds of my weekly budget unspent.

The two defects, both on the Claude Code surface

1. No consent, no estimate before a fan-out. When the agent decides to spawn dozens of subagents, nothing shows the order of magnitude beforehand and nothing shows the cost accumulating. A user cannot approve a spend they are not shown.

Ask: an estimate and a confirmation before agent-initiated parallel subagents, and a live counter while they run.

2. No warning before the wall. Nothing signals that the window is nearly full while a long task is in flight.

Ask: a warning at 75% and 90% of the window, and — since the agent knows the task ahead — a refusal to start a large fan-out that cannot finish inside the remaining budget.

Why it matters beyond one user

The product's argument is an agent that works on long horizons with autonomy. An agent that silently spends a user's quota and then dies mid-task inverts that argument: I now hesitate before letting it work unsupervised, which is exactly the behaviour the product is meant to remove.

I am a non-developer shipping a real product with Claude Code daily. Losing an hour mid-delivery costs far more than the tokens it saved.