XNews

OpenAI Codex Agent Harness Productized as Agents API

OpenAI Codex Agent Harness Productized as Agents API

OpenAI launched Agents API, productizing the harness used in Codex. Developers can create production-grade agents that run for days with one API call. The API handles context management, tool calling, and multi-agent orchestration; developers define agent capabilities. Execution environment options: self-hosted sandbox, partner sandbox (Blaxel, Cloudflare, etc.), or OpenAI-managed sandbox (new, based on Codex infrastructure). Key harness capabilities: long session compression near context limit, tool search (lazy loading), programmatic tool calling, subagent parallelism. An interview with Codex team lead Thibault reveals the harness always leads model improvements: tools are removed as models internalize capabilities. Codex merged into ChatGPT Work mode for $20/month plan.

Key insights

Agents API Addresses Production Bottlenecks

OpenAI's Agents API solves two common problems: harness (context management, tool orchestration, multi-agent) and infrastructure (reliable sandbox for days-long runs). The API runs agent loops on OpenAI infrastructure while developers retain control over capabilities and environment.

Flexible Execution Environments for Enterprise Adoption

OpenAI offers three sandbox options: self-hosted, partner (Blaxel, Cloudflare, DigitalOcean, etc.), and new OpenAI-managed sandbox. The managed sandbox supports file injection, packages, skills, and plugins, matching Codex capabilities.

Harness Includes Long-Running and Multi-Agent Features

Key features: automatic context compression for long sessions spanning multiple windows; tool search for on-demand tool definitions (saves tokens); programmatic tool calling for parallel/chain/filter; subagent parallelism where main agent delegates tasks to parallel subagents with independent contexts.

Codex Team Lead Reveals Harness Evolves Ahead of Models

Thibault explained that the harness (tools/scaffolding) always 'walks ahead' of models. The team builds crutches (e.g., /goal command) which later become internalized by next-gen models. The latest model no longer needs /goal; it can be told to work for a week.

Codex Merged into ChatGPT Work Mode

A major engineering challenge unified the locally-run Codex with cloud-based ChatGPT. ChatGPT Work mode now runs the full Codex harness in a cloud VM, capable of training models, running Blender, etc., included in $20/month Plus plan.

Open-Source and Model-Agnostic Strategy

Codex CLI and SDK are open-source and support non-OpenAI models. Thibault reasoned that if not open-source, competitors would fork with minimal changes. The downside includes competitors copying features before release and handling low-quality PRs.

Source posts