In the PR's cap-one live test, LiteLLM went from three searches, six Responses calls and a 500 to one search, two calls and a stopped 200.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
An administrator of LiteLLM, the AI gateway co-founded by Krrish Dholakia and Ishaan Jaffer, could set 'max_agentic_loops: 1', ask for three searches, and watch the gateway do the expensive part anyway. In the live test reported in the merged pull request, the request made three Tavily searches and six OpenAI Responses calls before returning HTTP 500. After the fix, the same cap-one request made one search and two Responses calls, returned 200 with 'finish_reason: stop', and logged $0.0312 in key spend instead of $0.0755.…
source & further reading
runtimewire.com — original article
llama.cpp patch speeds up one RTX 5090 matrix multiply 4.5x
Browser Use 1.0.5 sends unredacted screenshots to its default run judge
Hugging Face fixes TRL CISPO default that shrank gradients fivefold in its reproduction