OpenAI has made a strategic move to turn its coding agent Codex into a platform, launching the Agents API in public beta on September 10. The API allows developers to create agents by specifying just four parameters: task, model, tools, and runtime environment. OpenAI handles the underlying harness—the layer responsible for long-context compression, tool orchestration, and subagent collaboration—while developers can choose where the agent executes, using OpenAI's sandbox, their own infrastructure, or third-party environments like Cloudflare, E2B, and Modal. According to Chinese tech outlet TMTPost, OpenAI has been progressively dismantling Codex since April 2025, when it open-sourced the Codex CLI, followed by the cloud version in May, the SDK in October, and the App Server in February 2026. The Agents API now unifies these components under the banner of "Open Codex Harness," positioning Codex as a service rather than just a product.
The Battle for the Harness
The move highlights a growing competition over the "harness"—the infrastructure that enables AI agents to execute tasks reliably. DeepSeek, with its DeepSeek Harness (DSH), advocates for a modular, open ecosystem where every component, from models to tools to UI, is pluggable. In contrast, OpenAI's approach leans toward a managed cloud service, where the company hosts and maintains the harness, and developers simply pay for usage. Anthropic has been earlier in this direction, launching Claude Agent SDK in September 2025 and Claude Managed Agents in April 2026, which separate session, harness, and sandbox into independent layers. Google also entered the fray with Gemini API's Managed Agents at I/O in May. The competition reflects a shift from model intelligence to execution capability, where traditional platform companies with access to email, documents, and communication tools may gain an edge in the emerging "office agent" wars.
OpenAI's Agents API does not charge an additional platform fee; developers pay for model tokens and tools, and for compute if using OpenAI's sandbox. This pricing model could make it attractive for developers looking to build agent-based applications without managing infrastructure. However, the long-term winner will be the one whose harness becomes the default choice for developers, as the harness increasingly determines whether an agent can complete tasks end-to-end.
Based on reporting by 钛媒体. Edited and published in English by geisou.
