Jason Lee, a UC Berkeley academic, tested Claude Code in ultracode mode on large-scale coding setups and hit repeated failures where the agent claimed it needed more tokens and produced no output, even after breaking tasks into steps or setting max output to 64k tokens; this came after Ahmad Beirami recommended the tool or direct API for greater flexibility over chat limits.
@jasondeanlee @AnthropicAI Why don't you try Claude Code or the API?
API is much more configurable with your own harness
I paid more to @AnthropicAI to get more "continue". Will I ever get an answer?
Lee already ran the same large environments through Claude Code and saw identical token-related stalls, leaving open whether the direct Anthropic API would resolve the output failures or if the issue stems from session handling in resource-heavy ultracode workflows.
No official Anthropic response has appeared on these exact symptoms in large codebases, so it stays unclear how common the failures are beyond one researcher's ultracode and extra setups or whether recent tool updates altered the behavior.
No Digg Deeper questions have been answered for this story yet.
Even Claude code on ultracode doesn't work. It just keeps saying it needs more tokens and never outputs anything. I tried telling it to break it down to a smaller tasks etc and turning up the max output to 64k tokens. Any suggestions @AnthropicAI
Fable 5 on ultracode
@abeirami @AnthropicAI I tried Claude code also, same issue on ultracode and extra
@jasondeanlee @AnthropicAI Why don't you try Claude Code or the API?
API is much more configurable with your own harness