Responses API Adds Agent Computer Environment
๐กBuild secure, scalable agents with OpenAI's new Responses API runtime
โก 30-Second TL;DR
What Changed
Agent runtime built with Responses API, shell tool, hosted containers
Why It Matters
Empowers developers to deploy production-grade AI agents easily, accelerating agentic AI adoption in applications requiring persistent state and tools.
What To Do Next
Test the Responses API computer environment by building a simple file-handling agent.
๐ง Deep Insight
Web-grounded analysis with 6 cited sources.
๐ Enhanced Key Takeaways
- โขResponses API supports built-in tools including web search using the same engine as ChatGPT, file search, code interpreter, image generation, and remote MCP servers for external system connections[1][3][5].
- โขComputer use tool is powered by the Computer-Using Agent (CUA) model, achieving 38.1% on OSWorld, 58.1% on WebArena, and 87% on WebVoyager benchmarks, available as research preview for select developers[1].
- โขOpenAI introduced the open-source Agents SDK alongside Responses API for building multi-agent workflows with primitives like hosted tools, function tools, agents-as-tools, and MCP tools[3][4].
๐ Competitor Analysisโธ Show
| Feature | OpenAI Responses API | Azure AI Foundry |
|---|---|---|
| Built-in Tools | Web search, file search, computer use (CUA), code interpreter, image generation | CUA, function calling, file search |
| Integration | Agents SDK for multi-agent, stateful chaining | Azure AI Agent Service with Semantic Kernel, AutoGen |
| Pricing | Input/output pricing for computer use, batch API available, improved caching (40-80%) | Enterprise-grade, not specified |
| Benchmarks | CUA: 38.1% OSWorld, 58.1% WebArena, 87% WebVoyager | Not specified [1][2][5] |
๐ ๏ธ Technical Deep Dive
- โขComputer use tool processes screenshots as input and outputs actions like click, scroll, or type to control computers or virtual machines[1][4].
- โขAPI endpoint is /v1/responses, supports text/image inputs, JSON outputs, tool definitions, and state storage with store: true for multi-turn conversations[3][5].
- โขImproved performance with reasoning models like GPT-5 shows 3% better on SWE-bench; agentic loop handles multiple tool calls in one request[5].
๐ฎ Future ImplicationsAI analysis grounded in cited sources
โณ Timeline
๐ Sources (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- OpenAI โ New Tools for Building Agents
- azure.microsoft.com โ Announcing the Responses API and Computer Using Agent in Azure AI Foundry
- rahulkolekar.com โ How to Build Your First Production Ready Agent with Openai S Agents SDK and Responses API 2026 Guide
- community.openai.com โ 1140896
- developers.openai.com โ Migrate to Responses
- developers.openai.com โ Deprecations
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: OpenAI News โ