What did OpenAI announce on September 11, 2026, about the public beta of its Agents API—including how developers specify tasks, models, tools, and compute environments; OpenAI’s management of the orchestration harness; supported OpenAI-hosted, self-hosted, and partner infrastruct

OpenAI’s announcement, dated September 10 (reported September 11), put the Codex agent harness into public beta as a managed Agents API: developers define a task, model, tools, and execution environment, while OpenAI operates the agent loop and its ongoing improvements. 11

• Deployment and orchestration: Agents can run in OpenAI-hosted sandboxes, on a developer’s self-hosted infrastructure, or through integrated sandbox partners: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. OpenAI manages the harness, sessions, context, and recovery; developers retain control over tools, knowledge, and workloads. 11

• Agent capabilities: The managed harness automatically compacts context near the context-window limit; tool search loads relevant tool schemas only when needed; programmatic tool calling can parallelize and chain calls; and multi-agent mode delegates independent work to concurrently running subagents with separate contexts. 11

• Foundation, access, and price: The service is built on the open-source Codex harness, whose code developers can inspect, while OpenAI runs and maintains the managed version. It was made available in public beta to all developers with no Agents-API surcharge—users pay for model tokens and tools consumed—and OpenAI said it would iterate during beta toward general availability. 11

• Reported customer results: Ciridae reportedly improved an evaluation score from 0.71 to 0.85, while its CTO said concurrent subagents reduced latency by 4×. Those are customer-reported outcomes, not independent benchmark results. 6 8

• Simultaneous enterprise launch: OpenAI also introduced the Data agent for ChatGPT Work. It can analyze approved company data conversationally, investigate changes, produce shareable interactive dashboards, and use enterprise semantic context such as metric definitions and data relationships. 6

  • Supported data systems include Amazon Redshift, Datadog, BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake; it can also use Google Drive and SharePoint material. 6
  • It can operate across BI and collaboration systems—including Omni, Oracle BI, Power BI, Sigma, Tableau, ThoughtSpot, Slack, and email—subject to administrator-configured access and the underlying source permissions. 6

• Voice API: GPT‑Live‑1, OpenAI’s full-duplex voice model—designed to listen while speaking—also became available through the API, reported at $0.05 per minute. 4 12

• Yelp claim: I found insufficient reliable evidence in the available sources to substantiate a specific Yelp integration or a stated call-volume figure connected to this launch. It should not be treated as confirmed without a primary Yelp or OpenAI source.

• Competitive significance: OpenAI’s positioning is less “another agent framework” than a managed production runtime: it seeks to replace developer-built orchestration, context management, tool routing, durable sessions, sandboxing, and multi-agent coordination with an OpenAI-operated Codex harness. That is a direct practical challenge to frameworks in which developers assemble and run those layers themselves. This is an interpretation of the product design, not a claim that OpenAI named or benchmarked specific competitors. 11