Thanks for the update, @yekalkan. Excited to see the new release next week. Is there any update on my credits that were consumed by the breaking cache bugs?
Hey @yekalkan, any update on this ticket?
Thanks for the update, yekalkan. Looking forward to see this improve so we can fully engage the agent.
Hey yekalkan,
I think the issue is related to Claude's 'Prompt Caching'. Which is what you pointed out was the issue with my account. However, this is the first time the jeremygallas has ever used your agent. Is it possible that the bug/issue was still occurring with his account?
Let me know if you need more information.
[yekalkan] said: Hi again @harryhopkins
I am investigating the case. But it is not related with the previous issue, as i see.
Thanks. I agree that it probably is not related. I looked into the issue and discovered one of my engineers, jeremygallas, was performing a single session and used Opus on low reasoning. That is the session that used >$111 of credits, not mine. He was performing a test of your agent to review code that an OpenAI Codex (Sol) agent had implemented to compare the differences between the Codex agent and your agent. That was the only session they ran and it consumed that amount of credits in the session.
I was using your agent in a separate session when I saw the credit usage dramatically increase and thought it was my session. Thanks for investigating as we are working through our evaluation.
I don't think the issue has been resolved because it happened again today. But I think I can help provide more information to diagnose the issue. For instance, my current session today has used >$110 of credits, with $90 of credits being used in the last 10 minutes.
The issue appears when staying is the same session implementing multiple changes in phases or waves. Over time, the credit usage appears to accelerate. For instance, the first plan execution used about $13 of credits to analyze, plan, and execute. But each subsequent phase of analyze, plan, execute accelerates the credit usage.
Again, I want to continue using your AI agent because the workflow and code quality is better than frontier LLMs on their own. But I cannot keep losing $100 in credits by simply continuing a context in order to implement complicated features in phases vs. a single shot.
Your only response is to have me contact you again?
