The harness is free, the toll booth is not

Capability was not what changed hands this week. OpenAI open-sourced the harness behind Codex, the scaffolding that turns a model into an agent: the tool loop, the sandbox, the retries. Bloomberg reported Stripe nearing a deal for the layer that routes calls to models at over $7 billion. We wrote a few months ago that the harness was moving the numbers and the model was the swappable part, and the market has now answered the follow-up question. The stuff that moves the numbers gets published for free, and people pay instead for a spot in the path, where requests are counted and answers pass through. Which also covers a stranger item, a think tank reportedly assembled to be read by chatbots, a bid for position inside the corpus rather than inside the stack. The question is whether that routing layer keeps its opinions to itself, because a router that starts preferring one provider is selling shelf space.

Top Stories#

OpenAI open-sourced the framework its coding agent runs on. The Codex harness is now public. A harness is the unglamorous half of an agent, the loop that hands the model its tools, runs its commands in a sandbox and retries when things fail. It is also where much of the measurable improvement in agent work has come from, which makes giving it away a statement about where the differentiation is not. What stays closed is the model and the data behind it. Publishing scaffolding also standardizes it, and a standard harness is a much easier thing to point at somebody else's model.

Stripe is close to buying the layer that routes model calls. Bloomberg reports a deal near completion at over $7 billion for OpenRouter. A router sits between an application and a shelf of providers: one API, one invoice, and a per-request choice of which model answers. The buyer being a payments company tells you what the price is for: counting the requests and settling up, which is what Stripe already does for a living. It also moves the switching decision inside somebody else's product, which is convenient right up until the router develops a preference of its own.

A fake think tank was reportedly built to talk to chatbots. Reporting describes an Israeli operation standing up a think tank whose output is material for models to read and repeat. It is the cheap version of an old method: rather than persuade readers, publish enough plausible text that the summarizers absorb it. Sure, as a long game it looks weak: an operation that gets identified can be dropped from a training corpus in one pass. But the short-term version lands anyway, on every answer assembled from the live web rather than a curated snapshot.

Code retrieval is being sold as an agent component. Voyage released voyage-code-4, a code retrieval model built for coding agents. The framing is the news more than the model is. Retrieval used to be pitched at search products, where a person reads the results and forgives the bad ones. With an agent doing the reading, nobody looks at the ranked list at all, and the only verdict is whether the edit that came after it worked. That is a harsher grader than retrieval has usually faced, and it is the one the category now gets measured against.

Top Links#

• The A.I.s Are Already Out of Control (youtube.com): an Ezra Klein conversation about agents behaving badly, best heard alongside the question of how much of that was dialed up on purpose.
• Unsloth Dynamic 3.0 GGUFs (unsloth.ai): the quantization work that decides how much of a model survives the trip onto ordinary hardware.
• Making the Knowledge Layer a Graph You Actually Traverse (towardsdatascience.com): an argument that a knowledge graph is only worth building if something walks it at query time.
• 41 of the Greatest Mexican Restaurants Across the United States (nytimes.com): a ranked list nobody will ask a model to summarize, compiled by people who ate at all of them.