OpenAI says Codex SDK runs agents inside existing apps

• OpenAI is positioning Codex as an application-side agent runtime: the host keeps its UI, business context, tools, and approval flows while assigns the harness the agent loop.
• The integration splits into codex exec for bounded jobs, the SDK for programmatic task control, and app-server for persistent, UI-facing sessions, as describes it.
• A Codex tax-preparation pilot processed 7,000 returns and reduced reported preparation time by about one-third, per OpenAIDevs' tax-prep example, while gdb's post framed it as evidence the harness extends beyond coding.

OpenAI's own ARC-AGI-3 result says retained reasoning plus context compaction raised GPT-5.6 Sol from 13.3% to 38.3% while cutting output tokens sixfold. Its Relay sample puts the pattern in a fictional shipment console, with application-owned MCP tools and a human gate before rebooking freight.

OpenAI's platform post draws a deliberate boundary between the product and the harness:

The application owns

• The interface: dashboards, queues, editors, records, and approval flows.
• Domain context and business rules.
• MCP services, data, and actions.
• Execution location, accessible files and tools, observability, and approval requirements.

Codex owns

• Conversation state and work across turns.
• The agent loop, streamed execution, and tool interaction.
• Configured sandbox and approval-policy enforcement.

The released code is the harness and integration surface. OpenAI's post says model access and managed services remain separate.

The documented integration layers target different execution shapes:

• codex exec runs a bounded workflow for scripts, CI, or background jobs and returns structured output.
• The Codex SDK starts, resumes, and streams tasks from application code.
• Codex app-server connects a product to a local Codex process, keeps conversations open, streams events, interrupts work, exposes tools, and handles approval requests.

App-server exposes a client protocol for creating threads, starting turns, receiving events, and responding to approvals. The open-source Codex repository points builders to the Codex documentation for the surrounding CLI and integration surface.

OpenAI says Thrive Holdings and Crete used Codex in a tax-preparation workflow that incorporated practitioner feedback. The pilot processed 7,000 returns and cut preparation time by roughly one-third.

The same result appears in gdb's post, which casts it as a non-coding use for the harness. OpenAI's case-study paragraph does not publish a preparation-time baseline, filing-quality figures, or validation methodology.

Relay, OpenAI's sample app, makes the boundary concrete with a delayed-shipment workflow. A user selects a shipment and clicks Compare recovery; the application supplies context, the agent retrieves current sample operational data through application MCP tools, and then presents options.

A consequential write, rebooking the shipment, requires human approval. When the tool changes the record, the Relay description says the host application refreshes its business view, while the harness carries the conversation state and streamed activity underneath.