OpenAI Fully Open‑Sources Codex Harness: Inside the Agent Execution Framework

What Harness Is

Harness is an agent runtime that goes beyond a simple "model + prompt". It manages task understanding, context tracking, tool invocation, progress reporting, error handling, approval gating, and result delivery. The model generates ideas; Harness coordinates the execution pipeline.

Open‑Source Release

The source code is hosted at github.com/openai/codex under the Apache‑2.0 license. The repository is written in Rust, with most implementation in the codex‑rs directory. Three integration layers are provided, ordered by depth: codex exec: a command‑line tool for one‑off tasks (e.g., CI checks, background scripts). It runs a task to completion and returns a structured result.

Official SDKs in TypeScript and Python for launching, resuming, or streaming Codex tasks.

The app‑server, which exposes a JSON‑RPC protocol. Applications connect to a local Codex process, maintain long‑running conversations, receive event streams, interrupt execution, expose custom tools, and handle approval requests.

The official blog diagram shows the division of responsibilities: the application handles UI and business rules, the app‑server runs the agent loop and sandbox, and external MCP tools provide data and actions.

Performance Impact

On the ARC‑AGI‑3 benchmark, the same model (GPT‑5.6 Sol) with two Harness adjustments—retaining inference traces and applying context compression—raised the score from 13.3 % to 38.3 %, a near‑three‑fold increase. Output tokens dropped by a factor of six, reducing inference cost proportionally.

Real‑World Adoption

Cisco integrated the Codex SDK into its Cloud Control platform. Users describe requirements in natural language, and Harness generates custom applications.

Tax AI pilot (Thrive Holdings & Crete) processed 7,000 tax filings. Preparation time fell by roughly one‑third, accuracy peaked at 97 %, and throughput increased by about 50 %. Field‑completion rates rose from 25 % to 86 % within a month, with a projected 98 % by July.

GitHub and JetBrains embedded Codex into IDE workflows. JetBrains’ AI panel lets users select Codex as an agent and configure permissions ranging from simple Q&A to full network and command execution.

Relay demo showcases a logistics dashboard. A dispatcher selects a delayed shipment, the agent fetches real‑time data, computes recovery plans, and requires human approval before committing changes. The interaction occurs without a chat prompt; the UI itself supplies context.

Key Takeaways

The open‑sourced layer is the Harness and integration surface; model access and managed services remain separate. Providing a production‑validated agent runtime eliminates the need to build these components from scratch. The benchmark result—three‑fold score improvement with the same model—demonstrates that the execution system can have a larger impact than model upgrades.

Teams building internal AI‑augmented products can adopt Harness for ready‑made infrastructure such as agent loops, state management, sandboxing, and approvals, all under a permissive license. Even if Codex itself is not used, the layered architecture—keeping UI and business logic in the application while delegating the agent cycle to a framework—is a reusable pattern.

Project URL

https://github.com/openai/codex