OpenAI Deep Dive: How WebSockets in the Responses API Cut Codex Agent Loop Latency by 30-40%
OpenAI Blog·high signal
OpenAI published a technical deep dive on how WebSocket mode in the Responses API reduced overhead in the Codex agent loop. Persistent connections with connection-scoped caching eliminate per-turn context resending — for rollouts with 20+ tool calls, end-to-end execution is up to 40% faster. The blog details the architecture: only incremental inputs plus previous_response_id are sent each turn, with in-memory connection-local state enabling fast continuation. Critical reading for anyone building multi-step agent loops.