OpenAI stopped positioning Codex as a coding agent and started positioning it as a platform. On August 20, the company published Codex as a platform: build on the open agent harness, reframing its CLI, IDE extension, and SDK as one shared runtime — the same week DeepSeek’s open-source DSH (DeepSeek Harness) sailed past 173,000 GitHub stars. The AI competitive frontier is no longer just the model; it is the harness that runs the model, and the industry’s biggest model labs are suddenly fighting over it.
What OpenAI actually opened
The key move is unification. Codex App, the CLI, and the IDE extension all run on the same underlying harness, which manages session state, context, tool calls, sandboxing, and human approval. Developers can plug into that runtime through codex exec, the SDK, or app-server, and embed it into enterprise consoles, customer-service systems, security tools, and internal apps. OpenAI is effectively saying: stop thinking of Codex as a product, think of it as the execution layer underneath your products.
The numbers explain why the harness matters more than people assumed. On ARC-AGI-3, GPT-5.6 Sol running alone scored 13.3%. Wrapped in the Codex Harness, with continuous reasoning and context compression, the same model scored 38.3% — and cut output tokens to roughly one-sixth. The runtime is not a thin wrapper; it is a multiplier on model capability.
OpenAI also listed real deployments: Cisco uses the Codex SDK inside Cloud Control’s App Builder; GitHub and JetBrains have wired Codex into existing IDE workflows; Thrive Holdings and Crete used it for tax preparation, processing about 7,000 returns in a pilot while cutting preparation time by roughly a third. In each case, the pattern is the same: enterprise data, tools, approvals, and task logs all flow through one runtime that OpenAI supplies.
DeepSeek’s black whale: the runtime as a plugin box
Six days earlier, DeepSeek open-sourced DSH with a separate black-whale brand. Where Codex is a tuned engine you adapt around, DSH is designed as a box of interchangeable parts: model adapters, tools, skills, sessions, sandbox, storage, the agent loop, scheduling, and even the UI are all plugins, with a component called Cordis managing mount, unmount, and dependency relationships. In the car analogy circulating among developers: Codex gives you a well-tuned engine you can attach non-core components to, while DSH lets you assemble the engine itself.
The reception was unusual for an early preview. By August 20, DSH had roughly 173,700 stars and 18,800 forks; the launch thread hit 744 points on Hacker News. Chinese community testers built more than 200 plugins in two weeks, and the curated plugin list passed 270 entries within days of release, spawning Web UIs, mobile controls, model adapters, quota monitors, and plugin markets. The caveats are equally real: the codebase is about 453,000 lines across 219 packages, the first public commit was a single compressed squash, benchmark material is thin, and Reddit testers reported slow execution and high token consumption.
Why model companies must own the harness
A senior model developer I’d quote from the original reporting put it bluntly: beyond interaction-flow optimization, don’t touch the harness. Harnesses are hard to turn into a stable long-term moat — models change fast, and execution strategies differ per model; some plan first and call tools later, others correct as they go, and a tool set that one model juggles cleanly can destabilize another. By that logic, DSH cannot outrun Codex, because Codex sits on OpenAI’s models, and a harness without model strength is hollow.
So why build one at all? The answer is the closed loop. When DeepSeek only offered an API, its tasks ran inside external harnesses like Claude Code and Codex; DeepSeek saw the requests but not what happened before them or whether the task actually finished after the output. A harness lets a model lab observe both the model and the execution: where tasks fail, where tool calls stall, how many retries a task burns, which context strategy consumes the least tokens. That is training data for the next model and the next harness.
There is a business angle too. DeepSeek has no official token plan today; third-party platforms like OpenCode Go and Cola repackage its API into monthly subscriptions, owning the user entry point. With DSH, DeepSeek can measure sessions precisely — rounds per session, cache hit rates, wasted retry tokens, which steps should run on Flash versus Pro. A future managed DSH or subscription tier could shift billing from API tokens to quotas, task counts, concurrency, and Pro-model usage. Nothing is announced, but the instrumentation now exists.
The openness spectrum: Kimi Code, ZCode, and the rest
Chinese model companies were already building execution layers before DSH existed. Moonshot shipped Kimi CLI’s agent loop in September 2025 and evolved it into Kimi Code, an MIT-licensed terminal coding agent that manages context, tools, files, and commands — a complete harness, open for modification. Zhipu launched ZCode around GLM in July, an agentic development environment with its own agent managing tasks, context, terminal, files, permissions, and code review, but the core runtime stays closed; developers get plugins, skills, and MCP extensions, not the agent loop itself.
That creates a clear openness ranking: ZCode is GLM’s official execution environment; Kimi Code is an open coding agent and its runtime; DSH makes the disassemblable, recombinable runtime the product itself. Even Pi Harness, the MIT-licensed runtime whose maintainer is Flask creator Armin Ronacher, now matters again — Ronacher said DSH was the first new project in the space that made him reconsider some of Pi’s design choices. The race is not about stars; it is about which runtime becomes the default execution layer for real work, because the work itself generates the feedback that trains the next model.
What to watch
For teams, the harness is now an infrastructure decision, not a tool choice. DSH is an early preview with breaking changes ahead — fork it to learn, not to bet production on it yet, and demand benchmark material before committing. For builders, the runtime war decouples model choice from execution strategy: pick a harness by tool-call stability and token economics, not by which lab announced louder. And watch for two signals: whether DeepSeek ships a managed DSH or subscription pricing that moves billing off tokens, and whether Codex becomes the default agent base inside enterprises the way it already is inside Cisco and GitHub workflows. Whoever owns the runtime owns the loop — and the loop is where the next models come from.
For a hands-on walkthrough of DSH plugins, see DeepSeek Harness Plugins: Install, Market, Uninstall. For how Google is approaching the same layer from the training side, see EnvHarness: Agent Training Environments That Evolve.
FAQ
Q: What is the difference between Codex Harness and DeepSeek Harness?
A: Codex is a tuned engine you adapt around; DSH is a plugin box where you can reassemble the engine itself — model adapters, tools, sessions, sandbox, storage, the agent loop, scheduling, and UI are all swappable plugins managed by Cordis.
Q: Why did DeepSeek open-source its harness instead of keeping it internal?
A: The harness turns the runtime into a data loop: task failures, tool-call stalls, retry counts, and token costs flow back to the model team. DeepSeek sees not just API requests but what happens before and after them — feedback an API-only business never gets.
Q: Does the harness actually improve model performance?
A: On ARC-AGI-3, GPT-5.6 Sol alone scored 13.3%; inside the Codex Harness with continuous reasoning and context compression, the same model scored 38.3% while reducing output tokens to about one-sixth.
