OpenAI opened public beta access to the Agents API, a hosted service that packages the harness and infrastructure behind Codex and ChatGPT for Work into a single API call. A developer specifies a task, a model, a set of tools, and a compute environment, and OpenAI’s servers handle the orchestration: context management, tool routing, and subagent coordination that previously had to be built in house.
The product decision matters more than the feature list. OpenAI is not shipping a new model capability here. It is packaging the operational plumbing of running an agent for hours or days, the part most engineering teams find hardest and slowest to build, and renting it out as infrastructure. That reframes Codex less as a coding product and more as a chassis OpenAI can sell underneath anyone’s application.
Three specific capabilities anchor the pitch. Context compaction automatically compresses earlier turns of a long session as it approaches the model’s context limit, so a workflow can span multiple context windows without a developer writing their own summarization logic. Tool search loads only the tool definitions a task needs, cutting token overhead in agents wired to dozens of integrations. Programmatic tool calling lets an agent chain and filter tool results in code before anything reaches the model’s context, which is the difference between an agent that can search a database of ten thousand rows and one that chokes on it. A multi-agent mode lets a coordinating agent split a task and delegate pieces to subagents that run in parallel, each holding its own context.
Compute is where OpenAI is hedging. Rather than forcing every workload onto its own servers, the company is partnering with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel to run agent sandboxes on their infrastructure, alongside a fully OpenAI-hosted option for teams that want to skip the choice entirely. That is a wider sandbox bench than most rivals have assembled, and it signals OpenAI expects agent workloads to be heterogeneous enough that no single compute profile fits everyone.
Pricing is straightforward: no separate platform fee, just standard token and tool usage. OpenAI says the underlying harness is drawn from the open-source Codex codebase, which developers can inspect directly on GitHub, a transparency move that doubles as a trust argument for a black-box orchestration layer.
None of this exists in a vacuum. Anthropic already ships a comparable proposition through its Claude Agent SDK and Claude Code infrastructure, built on the same premise that the hard part of agents is the harness, not the model call. Amazon’s Bedrock Agents and Google’s Vertex AI Agent Builder make the identical bet from the cloud-provider side, and open frameworks like LangGraph compete on the assumption that developers would rather own that orchestration layer than rent it. The Agents API is OpenAI’s answer to all three at once: sell the harness, not just the model, and use Codex’s existing scale as proof it survives production load. The announcement includes a customer testimonial carousel but no named benchmarks or independent reliability data, so the durability claims here are OpenAI’s own until third parties test them at scale.
For any team currently maintaining a custom agent orchestration layer on top of the OpenAI API, the calculus is now a build-versus-rent decision with a real price on the “rent” side. Teams evaluating agent infrastructure over the next quarter should benchmark the Agents API against Anthropic’s Claude Agent SDK on the same workload before locking in either vendor’s harness.
According to OpenAI’s own announcement, “Introducing the Agents API,” published on the company’s site.
