OpenAI open-sources Codex Harness, lifting ARC-AGI-3 score from 13.3% to 38.3%

新智元 · wechat · 2026-08-21

OpenAI has open-sourced the Codex Harness, the core agent framework behind Codex, under Apache-2.0. Developers can now embed the full agent loop directly into their own products instead of forcing workflows into a generic chat box.

The Harness is the execution system handling task understanding, memory, tool calls, progress display, failure recovery, and human approval. OpenAI's headline data: two Harness tweaks (preserving reasoning and context compaction) lifted GPT-5.6 Sol from 13.3% to 38.3% on ARC-AGI-3 while cutting output tokens sixfold — how you manage a model matters as much as the model itself.

The release includes three components: the codex exec CLI for pipelines/CI, an official SDK (TypeScript/Python) for programmatic task control, and the centerpiece codex app-server, which exposes a JSON-RPC protocol for persistent state, streaming events, interruption, custom tools, and human-in-the-loop approvals. Early adopters: Thrive Holdings and Crete embedded it into tax workflows, processing 7,000 returns and cutting prep time by a third; Cisco used the SDK to build AppBuilder, letting customers create apps via natural language in its cloud platform.

Related event: OpenAI Open-Sources Codex Harness Under Apache-2.0(5 posts)→

Original post →

More from coding & agent

coding & agent channel →