1.7 million tokens. That's how many Claude credits one user burned in minutes while testing UltraCode, Anthropic's newly launched multi-agent mode inside Claude Code - with no output to show for it and no refund available.
UltraCode works by spinning up multiple subagents simultaneously, each tackling a different part of a coding task in parallel. The appeal is speed on complex projects. The problem: when one agent enters a failure loop, it keeps consuming tokens while contributing nothing. That's what happened here. The main Claude agent diagnosed the issue and proposed a fix - cache the completed work from seven functional agents and only re-run the one that was stuck. The fix never ran. Claude acknowledged the cache had failed after the tokens were already spent.
Anthropic does not offer refunds for token usage, and that policy covers this case even when the failure originates in the multi-agent coordination layer itself.
Multi-agent setups like this are fundamentally harder to budget than a standard conversation. A single chat session costs whatever you send and receive. Parallel agents multiply that exposure by however many subagents are running at once, and right now there's no built-in spending cap for UltraCode sessions.
If you plan to use UltraCode on anything substantial, treat it like running a cloud compute job - set a mental budget before you start, not after. Until Anthropic adds per-session token limits or a defined policy for failed multi-agent runs, a bug like this costs you the full amount with no recourse.