OpenAI says its own agent rewrote the kernels that serve its models, and publishes the harness rules alongside

The claim practitioners have been passing around all week has a first-party source. OpenAI writes that "With Codex, GPT-5.6 Sol autonomously rewrote and optimized our production kernels, the core code that executes the mathematical operations that make up the model," and that this and related work "reduced end-to-end serving costs by 20%." A second agent-run programme on the speculative decoding draft model, "designing and running hundreds of experiments on its architecture," is credited with raising token generation efficiency by more than 15%. The companion pricing post frames all of it as happening "Within a human-led process," which is the hedge the secondhand versions of this story have been dropping.

OpenAI says it has "heavily invested in verification tooling, such as the open-source tool FpSan (Floating-Point Sanitizer), to help validate the correctness of the kernels written by GPT-5.6 Sol." The post also describes the Codex harness in concrete terms. Tool output is "capped at 10,000 tokens by default unless the model requests a different limit." Integrations, MCP tools, skills and plugins are surfaced only on demand. And model-visible history is treated as "append-only: new messages, tool results, and environment updates are added at the end rather than inserted into earlier context," so the cached prompt prefix survives the turn.

Why it matters: The kernel story will get the attention, but the harness section is the part you can use tomorrow. Append-only context, deferred tool discovery and a hard cap on tool output are three rules you can put in your own loop this week, published by a team running one at scale. Treat the autonomy claim as an unaudited vendor claim about its own internal work, because that is what it is.