Open Reasoning Standard for Agents Unveiled

As AI workflows shift toward agent-centric approaches, Open Responses, a new open reasoning standard designed to overcome the limitations of the existing Chat Completion format, has been announced. This standard is based on OpenAI's Responses API and was built as open source with support from the Hugging Face ecosystem.

The key features of Open Responses are as follows:

• Stateless by default: It does not maintain state by default, while supporting encrypted reasoning when needed.
• Standardized model configuration: Model configuration parameters are standardized to improve interoperability.
• Semantic event streaming: Rather than simple text or object deltas, streaming is modeled as a series of semantic events.
• Extensibility: Configuration parameters specific to particular model providers can be used.

Through this, developers can secure richer reasoning visibility. Whereas previously only summarized information was provided, now the model's thought process can be handled in greater detail through three fields: raw reasoning traces, encrypted content, and summary. It also supports detailed state changes and payloads, such as the code interpreter's execution status (e.g., interpreting), enabling more precise monitoring of agent behavior.