The chatbot era is dead: Open Responses aims to fix the agent mess

The Chat Completion API is a zombie. It’s still shuffling around as the de facto standard for AI inference, but it was built for simple back-and-forth conversations, not the autonomous, multi-step agentic workflows that now dominate. That mismatch is creating real friction. The new Open Responses specification, an open-source extension of OpenAI’s Responses API launched in March 2025, is a direct move to put it out of its misery. The goal is to give developers and inference providers a shared, practical format for systems that don’t just talk, but reason, plan, and act.

The core of the problem is state and visibility. Agentic tasks aren’t single turns. They’re loops that spin off tool calls and run for unpredictable timeframes. The Responses format was designed for this, supporting things like provider-side agentic loops and task-based endpoints for video. Open Responses builds on this by being stateless by default and, crucially, standardizing how reasoning is exposed. It formalizes three fields for reasoning items: content for raw reasoning traces, encrypted_content for protected provider-specific data, and summary for sanitized versions. This means clients migrating from providers like OpenAI, which historically only offered summaries and encrypted content, can now potentially tap into raw reasoning streams from other model providers.

Streaming gets a major overhaul too. It’s no longer about just raw text deltas. Open Responses models streaming as a series of semantic events. For example, you’ll see specific events like response.reasoning.delta for raw reasoning chunks from open-weight models, and response.reasoning_summary_text.delta for summaries. A hosted Code Interpreter can even fire an interpreting state event to give users visibility during long operations. It’s a shift from seeing the model’s raw output to observing its cognitive process and actions.

This is a play for interoperability. The specification includes standardized model configuration and extensible parameters for specific providers, which is a clear signal to routing providers that they can standardize on a consistent endpoint. The migration path is designed to be low-friction for clients already using the Responses API. The big question is adoption. The Chat Completion format is deeply entrenched, and OpenAI’s own Responses API is still closed. An open standard only works if the ecosystem rallies around it, and that’s never a guarantee. The coming months of collaboration with the community and providers will determine if this becomes the new normal or just another well-intentioned fork in the road.