On September 10 2026, OpenAI released the public beta of its Agents API, a managed cloud service that lets developers create agents using the same harness as Codex, specify models, tools, and compute environments, and automatically handle long-session context, sub-agent parallelization, and sandbox orchestration.
Short answer: On September 10 2026, OpenAI released the public beta of its Agents API, a managed cloud service that lets developers create agents using the same harness as Codex, specify models, tools, and compute environments, and automatically handle long-session context, sub-agent parallelization, and sandbox orchestration.
On September 10, 2026, OpenAI rolled out the public beta of the Agents API, a managed service that lets developers spin up cloud-based agents using the very same harness that drives Codex. The idea is to take the pain out of building agents that can chug along for hours, call tools, and orchestrate sub-agents without having to stand up and maintain their own orchestration layer.
Getting started is as simple as firing off a single API call where you spell out the agent’s task, pick a model, enumerate the tools it may use, and point to a compute environment. In the sample request the model is set to “gpt-6-astra,” a tool of type MCP points at an observability endpoint, multi-agent mode is switched on with a ceiling of three concurrent sub-agents, a vault ID is supplied for secure storage, and an OpenAI-hosted sandbox is configured with a specific capabilities directory. The input string walks through a scenario where the agent spots a surge in 5xx errors, farms out the analysis to sub-agents, and stashes the findings in a workspace folder.
OpenAI takes care of the underlying harness and the sandbox plumbing, but you’re not locked into one way of running things. Agents can execute in OpenAI-provided sandboxes, on your own hardware, or inside sandboxes offered by a growing roster of partners-including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. This smorgasbord of options lets teams line up the agent’s compute, storage, and deployment needs with their existing workflows while still riding a harness that gets refreshed in lockstep with new model releases.
One of the headline upgrades in the harness is how it handles context over long hauls. As an agent bumps up against its context ceiling, the API automatically compacts earlier exchanges, keeping the crucial bits alive for continued operation. That spares developers from writing their own compaction logic and opens the door to workflows that stretch across multiple context windows-think multi-hour investigations or day-long data crunches. The harness also nudges agents to use their tools more judiciously, trimming needless calls and boosting overall efficiency.
The Agents API bakes in native support for sub-agent parallelization. Flip on multi-agent mode and the service spins up several workers to tackle different slices of a problem, gathers their results, and stitches them into a final answer. Early adopters have touted the payoff: noticeable latency drops and the ability to fan out work across hundreds of agents during peak traffic without keeping idle servers humming in the background.
Feedback from the beta crowd paints a vivid picture of the impact. One CTO saw their evaluation score jump from 0.71 to 0.85 after adopting the API, chalking the improvement up to sub-agent support that sliced latency by four times. An engineering leader reported a sixty percent dip in cost per case, alongside lower latency and better token efficiency, all while performance held steady. A co-founder described how the API swallowed bursty loads by letting hundreds of agents run asynchronously, doing away with the need for idle capacity between spikes. A lead engineer in financial services noted an eighty-six percent plunge in failed agent responses after the harness was decoupled from the sandbox, which gave a big confidence boost for production rollouts. Other teams highlighted how the API unlocked complex, multi-step workflows that previously demanded custom orchestration code.
If you’re keen to give it a spin, the public beta lives right on the OpenAI platform. The documentation packs quick-start guides, sample code, and details on configuring sandboxes or hooking into partner environments. By playing around with the API, teams can gauge how managed harnessing, long-session context handling, and sub-agent orchestration might streamline their own AI-driven pipelines and shave off the engineering overhead that usually comes with running reliable agents at scale.
On September 10, 2026, OpenAI released the public beta of the Agents API, a managed service that lets developers create cloud-based agents using the same harness that powers Codex, handling orchestration, tool use, and sub-agent coordination without needing to build their own infrastructure.
Agents can run in OpenAI-hosted sandboxes, on a developer’s own hardware, or inside partner-provided sandboxes from Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, or Vercel, letting teams match their existing workflows while staying on a harness that updates with new model releases.
When an agent reaches its context limit, the harness automatically compacts earlier exchanges, preserving essential information so developers don’t need to write their own compaction logic; it also encourages judicious tool use, trimming unnecessary calls and boosting overall efficiency for multi-hour or day-long workflows.
Early adopters saw evaluation scores rise from 0.71 to 0.85, latency cut four-fold, cost per case drop sixty percent, token efficiency improve, failed agent responses fall eighty-six percent after harness-sandbox decoupling, and the ability to run hundreds of agents asynchronously to handle bursty loads without idle servers.
Source: OpenAI
On September 10, 2026, OpenAI launched ChatGPT for Financial Services, a product that integrates built-in financial data with the GPT-6 Astra architecture to support research, modeling, and client-ready material creation within the chat interface, according to the announcement.
Microsoft announced on September 9, 2026 that it has agreed to a set of AI safety and privacy principles for use in US schools, pledging not to train its models on student or educator data, limiting data collection, providing plain-language explanations to families, banning AI companions that interact directly with learners, and requiring human review for high-risk decisions, with districts able to opt-in starting in November.
On September 9, 2026, OpenAI announced that its unreleased AI model solved the Navier-Stokes Millennium Prize problem in 88 hours using a swarm of about ten thousand internal agents, a claim that sparked controversy after NYU professor Tristan Buckmaster alleged the company pressured him to drop co-author Levent Alpöge and questioned its use of his data.
NO COMMENTS YET
Comments are open. Have a thought or a question? Share it below.