OpenAI’s Agents API reportedly launches a public beta

• OpenAI has opened the Agents API to all developers in public beta, and minchoi's launch roundup describes its basic request as a task, model, tools, and execution environment.
• Durable sessions can carry work across context windows, with automatic compaction and parallel subagents listed in minchoi's feature inventory.
• Developers can choose OpenAI-hosted, partner-hosted, or self-hosted compute, a division shown in minchoi's architecture diagram.
• The API has no separate Agents fee, as minchoi's pricing summary notes, but OpenAI's documentation separately bills hosted sandbox containers alongside model and tool usage.
• MCP is already reaching visual workflows: magnific's ChatGPT post places a motion-graphics workflow in a conversation where camera, pacing, and graphic effects can be revised together.

The official API overview defines a session as a durable agent instance that can be steered while it works. Its hosted-sandbox guide describes a Linux workspace with Python, Node.js, and command-line tools, while magnific's signal-film brief turns a 30-second animation into a shot-by-shot specification.

OpenAI's launch announcement says it operates the Codex harness, including orchestration, context management, and recovery. The application supplies the tools and selects the environment where the work happens.

The API overview breaks the system into four objects:

• Agent: model, instructions, tools, and MCP servers.
• Environment: an optional sandbox or computer for files, skills, and commands.
• Session: the durable instance that performs work and accepts follow-up input.
• Events and items: the application's inputs and the agent's output during that session.

OpenAI says the core Codex harness is open source, while the API provides its operated version and the surrounding cloud infrastructure.

A session retains the agent configuration, conversation, and saved work, according to OpenAI's sessions documentation. The application can create it, wait for the environment, send a task, follow progress through a stream or webhooks, then continue or steer the same run.

As a session approaches its context limit, the launch announcement says the harness automatically compacts earlier context while preserving material needed for the work ahead. That removes the need to build a separate compaction loop for workflows that span multiple windows.

The hosted-sandbox guide says OpenAI-hosted environments provide a Linux workspace with Python, Node.js, and command-line tools. Packages and input files are prepared before setup commands run; a failing setup command prevents the agent from starting.

OpenAI's launch post offers three execution paths:

• OpenAI-hosted: OpenAI provisions and manages the sandbox.
• Sandbox partners: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.
• Self-hosted: the application controls its own infrastructure, including a custom image, compute, or private network.

The API supports MCP servers, custom functions, and built-in tools such as web search, per OpenAI's launch announcement. Its tool layer has two separate optimizations:

• Tool search loads definitions only when relevant, which OpenAI says can preserve cache and reduce token use.
• Programmatic tool calling lets an agent run calls in parallel, chain operations, and filter or combine results in code before returning only relevant material to context.

The API overview also lists artifact production, file editing, and live steering as managed-harness capabilities.

For independent parts of a task, the API can split work across subagents in parallel. OpenAI says each subagent gets its own context while the primary agent coordinates the results, with concurrency configured through multi_agent.

LukeW's posts describe how Intent, a separate agent-coordination product, organizes larger fleets:

• Each task gets a dedicated workspace copy and a living spec, according to LukeW's focused-context post.
• A coordinator delegates work, implementers write it, and verifiers check acceptance criteria in LukeW's agent-roles post.
• Hooks wake an agent when a meaningful event occurs, such as a pull-request review comment, in LukeW's handoff post.
• A finished agent briefs a fresh one and retires its own subagents in LukeW's succession post.
• Intent can distribute agents across workspaces and devices, LukeW's multi-device post says.

OpenAI says there is no additional fee for the Agents API itself. Its API overview prices model usage at the selected model's API rate, OpenAI tools at their standard rates, and OpenAI-hosted sandboxes at standard container rates.

The same documentation currently limits data residency to the United States and says the API does not support Zero Data Retention. Choosing a self-hosted sandbox does not change that status.

gregisenberg argued in a follow-up to gregisenberg's Agents API post that rented orchestration shifts differentiation toward the parts of a workflow a general harness does not contain: proprietary tools and data, approvals, and the definition of a completed job.

The jobs he named were:

• Freight exceptions
• Insurance reviews
• Security triage
• Revenue leakage
• Healthcare administration
• Compliance operations

magnific says in magnific's ChatGPT post that its MCP workflow can revise camera, pacing, and graphics in the same ChatGPT conversation.

The 30-second brief in magnific's signal-film prompt separates the film into six controllable layers:

• A reference used for editing energy and compositing, with explicit instructions not to reproduce its imagery or audio.
• Location, palette, materials, and visual language.
• One visible causal path, an orange signal moving from a wall junction box to a rooftop antenna.
• Twelve timed shots, generally two to three seconds each.
• Interaction rules for scan grids, paper-cut geometry, metal slats, and a moire print.
• Music, cut rhythm, framing, and exclusions such as logos, typography, and generic futuristic forms.

A related magnific character-film prompt pins one courier's appearance and a single orange audio cartridge across a 15-second sequence, then specifies when the cartridge leaves her hand and activates the world.