OpenAI is responsible for sessions, orchestration, context compaction, and recovery. An application gets access to the Codex harness through the OpenAI API, while the developer supplies tools and chooses the execution environment.

Agents can work in a sandbox: execute code, edit files, connect to MCP servers — bridges to external tools — and produce work outputs. Model usage is billed at the selected model’s rates, while OpenAI tools and sandboxes use standard rates.

Claim check:

  • The Agents API offers a managed Codex harness for cloud agents. (confirmed by the publication itself: evidence; «Build durable cloud agents with a managed Codex harness.»)
  • The Agents API gives an application access to the Codex harness through an OpenAI-managed API. (confirmed by the publication itself: evidence; «The Agents API gives your application access to the Codex harness through an OpenAI-managed API.»)
  • OpenAI manages sessions, orchestration, context compaction, and recovery. (confirmed by the publication itself: evidence; «OpenAI manages sessions, orchestration, context compaction, and recovery while your application provides tools and chooses its execution environment.»)
  • The developer supplies tools and chooses the agent’s execution environment. (confirmed by the publication itself: evidence; «OpenAI manages sessions, orchestration, context compaction, and recovery while your application provides tools and chooses its execution environment.»)
  • Agents can execute code, edit files, connect to MCP servers, and produce artifacts in a sandbox. (confirmed by the publication itself: evidence; «Agents can operate in a sandbox where they can execute code, edit files, connect to MCP servers, and produce artifacts.»)
  • Model usage is billed at the selected model’s rates, while OpenAI tools and sandboxes use standard rates. (confirmed by the publication itself: evidence; «Model usage is billed at the selected model’s API rates . OpenAI tools use their standard rates , and OpenAI-hosted sandboxes use standard container rates .»)

Primary sources:

score 81.5 out of 100 · kind: announcement