OpenAI Agents API: Managed Codex Harness as a Product

On September 10, 2026, OpenAI released the Agents API in public beta, exposing the same harness that powers Codex and ChatGPT — session orchestration, context management, tool coordination, subagent spawning, error recovery, and sandboxed execution — as a general-purpose managed service. The one-line pitch from OpenAI's announcement: "Build and run cloud agents with the Codex harness, fully managed by OpenAI."

The structural signal is clear: intelligence is commoditizing, and value is migrating up the stack from the model to the harness that turns a model into a reliable process. The Agents API is OpenAI's bid to own that layer.

• Managed Codex harness: OpenAI operates and maintains the session loop, context compaction, tool search, programmatic tool calling, and parallel subagents — the same machinery that runs Codex
• Three execution environments: OpenAI-managed sandbox, your own infrastructure, or one of 9 partner sandboxes (including Vercel, Oracle, Cloudflare)
• Persistent sessions: Agents run for hours or days with saved turns, items, and intermediate state
• Automatic context compaction: The harness manages context window overflow so you don't hand-roll it
• Subagent coordination: Built-in parallel subagent spawning and coordination
• Open-source harness codebase: The Codex harness is open source, so you can inspect the core logic even though OpenAI runs it for you

• You want a reliable long-running agent without building and maintaining your own harness
• You need managed sandboxes with code execution, file handling, and artifact production
• You want to skip the orchestration-layer framework tax (LangChain, CrewAI, etc.) and use the harness that already runs at OpenAI's scale
• You're migrating from the sunset Assistants API (deprecated August 26, 2026)

• You need full control over the harness logic and don't want a managed dependency
• You need data residency outside the US (current beta is US-only data residency)
• You need Zero Data Retention for regulated workloads (not yet supported in beta)
• You want to use non-OpenAI models as the agent's brain (the harness is OpenAI-ecosystem)

The API is organized around 4 core concepts: Agent (configuration), Environment (compute/sandbox), Session (state), and the harness loop that ties them together. There is no extra fee for the harness itself — you pay for tokens, tools, and container time.

• No harness surcharge: You pay for model tokens, tool usage, and container/sandbox time only
• Token pricing follows standard OpenAI model rates
• Container time billed for sandbox compute usage

• Claude Agent SDK — Anthropic's equivalent; harness that powers Claude Code, Python/TypeScript SDK, bills against Claude subscription plans
• Google ADK 2.0 — Google's Agent Development Kit with Antigravity CLI and Managed Agents API
• LangGraph / CrewAI — self-managed orchestration frameworks; more control but you own the infrastructure
• Temporal — durable execution framework for long-running workflows; not AI-native but battle-tested for session management