Open Responses Joins AAIF with AWS Support

The API format that a growing number of model providers use to call language models now has a vendor-neutral home. Open Responses has joined the Agentic AI Foundation (AAIF) as a project, bringing it under open governance alongside MCP and A2A.

Based on the OpenAI Open Responses API, Open Responses is an open source specification that defines a single API format for calling language models, streaming results, and orchestrating tool use. Build your agent against an open API format and keep the freedom to evaluate new models, switch providers, or run multi-provider setups without reworking your application logic.

AWS is involved in Open Responses and employs a core maintainer on the project, and Amazon Bedrock already supports the Responses API across dozens of models. You can start building today against an API format that’s on a path to becoming an industry standard.

What Open Responses is

Developers building agentic applications face a fragmented landscape of model APIs. OpenAI pioneered the Responses API for stateful conversations that call external tools or functions. Anthropic developed its own Messages API with different conventions for tool use and streaming. Google’s Gemini API takes yet another approach.

Each encodes similar primitives including messages, function calls, tool usage, and multimodal inputs but does so differently enough that your orchestration code is tied to whichever provider you started with. Open Responses standardizes these concepts into a shared specification with versioning, compliance tests, and extensibility rules, that defines:

• A common request/response model — one schema for structured inputs and outputs that works across OpenAI, Anthropic, Gemini, local models, and any provider that implements the spec.
• An agentic loop — a unified pattern for models to emit tool calls, receive results, and continue reasoning, whether tools are developer-hosted or provider-hosted.
• Items as the atomic unit of context — a clear schema for messages, tool invocations, and reasoning state, with defined state machines (in_progress, completed, failed) for predictable behavior.
• Semantic streaming events — meaningful event types like response.output_text.delta and response.completed rather than raw text deltas, so clients can be built once and work across providers.
• Extensibility without fragmentation — provider-specific tools and item types are namespaced with provider slugs, keeping the portable core clean while preserving room for innovation.

Practically, this means you pick a model provider based on which model is best for the job, not which SDK you’ve already committed to. Want to evaluate Claude for reasoning and GPT for code generation? A unified request/response format lets you run both through the same orchestration code. Need to switch providers six months from now? Your orchestration logic doesn’t change.

AWS supports Open Responses

At AWS, we believe developers should be able to choose the best model for each task without that choice dictating their entire architecture. By supporting open source and open standards we give you the freedom to compose, extend, and migrate your workloads on your own terms.

The agentic AI ecosystem is growing fast, and with it, the number of decisions that can lock you in, including which SDK you use to call a model, how your agents discover tools, how they talk to each other, and how you handle streaming and conversation state. Each of these choices can become a switching cost if it’s tied to a single vendor’s conventions.

We’re backing Open Responses with both engineering contributions and production support across AWS services.

Amazon Bedrock has supported the Responses API since December 2025 and continues to expand its implementation. Developers can use the familiar OpenAI SDK to call models hosted on Bedrock just by changing the base URL and API key. If you’re already using Amazon Bedrock’s Responses API, you’re already building on this API format. Bedrock’s Responses API support includes:

• Stateful conversation management via previous_response_id. No need to pass the full conversation history with each request.
• Built-in tool use for agentic workflows including developer-hosted and server-side custom tools, with server-side tool support added in January 2026.
• Streaming and non-streaming modes across both bedrock-runtime and bedrock-mantle endpoints.
• Broad model availability as the Responses API works with models from Amazon, Anthropic, Meta, Mistral, OpenAI, DeepSeek, Google, and others, all through the same API format.

This means existing applications built against the OpenAI Responses API can run on Bedrock with minimal changes, gaining access to Bedrock’s features such as Guardrails, cross-Region inference, and intelligent prompt routing, while keeping the same client code.

AWS is a member of the AAIF Governing Board and is contributing to Open Responses’ development as an AAIF project. By housing Open Responses under the AAIF alongside MCP and A2A, we ensure that the specification evolves through community consensus, not any single company’s product roadmap.

As the specification matures under the AAIF, your existing code aligns more closely with the formal standard. It’s straightforward to add a second provider or move a workload without starting over. Bedrock’s implementation also provides a production-grade reference for how the spec works at scale.

If you’re a model provider or building an inference platform, Open Responses gives you a clear specification to implement, with one schema and compliance tests included, so your users can integrate without learning yet another proprietary API.

Looking ahead

Open Responses joining the AAIF is an important step, but the specification continues to evolve. Recent additions include WebSocket transport for persistent connections and a compaction endpoint for managing long conversation contexts. The compliance testing suite, available both as a web UI and a CLI, ensures implementations stay aligned as the spec grows, and open-weight serving is spreading, with Ollama, vLLM, OpenRouter, and Hugging Face now among the providers serving the format.

Together with MCP, A2A, Agent Plugins, and now Open Responses we’re building toward an agentic AI ecosystem where every layer of the stack is portable, composable, and open.

Explore the Open Responses specification at openresponses.org, the GitHub repository, and get started with the Responses API on Amazon Bedrock.

———

Anthony Liguori is a VP and Distinguished Engineer, Libby Clark is a Principal Open Source Strategist and James Ward is a Principal Developer Advocate at AWS & member of the AAIF Technical Committee.