OpenAI’s Codex Harness Claim Is Unverified—but Context Management Tripled ARC-AGI-3 Scores

These sources support a practical picture of the system: an execution runtime around the model, with tools, state, feedback, and interfaces that let an agent perform multi-step work. They do not, on their own, establish that OpenAI made a newly named “Harness” available under the Apache 2.0 license on August 20.

What is not verified

The supplied evidence does not establish:

• that an August 20 release occurred;
• that the release was formally named “Harness”;
• that it was distributed under the Apache 2.0 license;
• which exact components were included in that alleged package; or
• which companies had adopted it and what results they achieved.

The claim appears in a secondary report that describes

codex exec

Why retained reasoning and compaction mattered on ARC-AGI-3

A separate OpenAI experiment provides the clearest source-backed result in the material. Using GPT-5.6 Sol on the ARC-AGI-3 public task set, OpenAI reported a score of 13.3% with the standard harness. After enabling retained reasoning and context compaction through the Responses API, the score reached 38.3%—roughly a threefold improvement—while output-token use fell by about six times .

The model weights were not the change described in that comparison. The important differences were how the runtime carried state between turns:

• Retained reasoning allowed useful intermediate reasoning to persist across actions instead of forcing the model to reconstruct its approach after each step.
• Context compaction compressed older task history rather than simply allowing earlier context to disappear as the interaction grew.

OpenAI’s builder guidance presents these settings as ways to reuse work already performed and maintain coherence over longer task horizons . The result is therefore best understood as a measurement of the combined model-and-runtime system, not as evidence that the underlying model alone became three times more capable.

The benchmark result also comes with an important limitation: it compares two execution configurations on a public evaluation set. It does not prove that the same improvement will appear in every agent, model, workload, or production environment.

What developers can use today

The strongest confirmed developer takeaway is not the unverified Apache 2.0 announcement. It is that OpenAI provides several ways to build around agentic coding behavior:

• Direct API access: the Codex prompting guide positions API use as the most customizable option .
• Codex SDK: the SDK is presented as a simpler integration path for developers who do not want to assemble every interface themselves .
• Open-source CLI reference: OpenAI identifies the Codex CLI as a reference implementation that developers can inspect and adapt .
• Agent execution loop: the documented plan–implement–validate–repair cycle provides the operating pattern behind long-horizon coding work .
• Cloud execution: Codex can work with repositories, run code, and handle background tasks in parallel in its cloud environment .

Together, these capabilities make it possible to place an agent inside a product or engineering workflow instead of limiting it to a general-purpose conversational interface. Developers can supply domain-specific instructions, tools, permissions, validation steps, and state handling around the model. That is a documented integration direction; the broader claim that a single Apache-licensed Harness release formalizes all of it remains unconfirmed by the supplied official sources.

Adoption evidence is narrower than the release claim

The supplied sources do not identify companies that adopted the alleged Apache-licensed framework, nor do they provide company-level outcomes for legal, operational, research, or other non-coding workflows.

There is broader evidence that Codex is being used by enterprise customers: OpenAI reported that, as of June, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers . That figure indicates substantial enterprise activity, but it does not demonstrate adoption of the alleged Harness package, identify particular companies, or establish business impact.

The practical strategic implication

The verified pattern points toward a shift from chat-based assistance to delegated execution. Codex combines a model with tools, an iterative feedback loop, persistent or compressed state, and execution environments capable of carrying out work over multiple steps .

For product teams, that architecture can support specialized agents for software repositories, internal operations, research tasks, or other controlled workflows. The quality of those applications will depend not only on model choice, but also on how developers design tool access, state retention, context limits, validation, and recovery from failure.

That is the durable lesson from the ARC-AGI-3 experiment: changing the harness can change both capability and efficiency. But based on the evidence available here, it is more accurate to describe OpenAI’s Codex runtime and API components than to report the August 20 Apache 2.0 “Harness” release as an established fact.