The September 3 edition looked at nine harnesses running one model with a 17x spread in cost, and treated the wrapper around the weights as a measurement problem. This week the wrapper acquired vendors. NVIDIA Labs released SoL-Pi, a standalone extension to the Pi agent harness that wraps AutoResearch-style loops around the agent to cut token spend without cutting useful work. PI-Desktop shipped an Electron shell over a Rust host core running the same harness, with user-installable plugins. Tencent put out teamai-cli, pitched at making a whole team AI native rather than accelerating one developer. Three artifacts, one day, all of them selling scaffolding rather than weights.
The versioning is the part that matters. A harness with its own release notes, its own plugin host and its own upgrade cadence is a component that can regress independently of the model underneath it. Attribution gets harder in both directions. A win after a model swap may belong to the harness upgrade that landed the same week, and a regression may belong to neither.
Traffic runs the other way too. Cognition launched SWE-2, a coding-specialized model from a company whose product to date has been the harness. The parity claims against Fable 5.1 and GPT-Astra are marketing until someone runs them against their own repository, but the availability is real, and it points at a market where the harness vendor and the model vendor are the same company. Lock-in arguments that assumed the model was the sticky part get less clean.
The provenance thread the August 31 edition opened, on who gets to say what shipped, picked up a second accuser. A researcher claims OpenAI trained on their chat transcripts and then announced the result as its own breakthrough. The merits are unresolved and may stay that way. The operational fact underneath is settled and has not moved all week: unpublished work pasted into a hosted agent has been disclosed, whatever the terms of service say about it.
John D. Cook wrote up the detail most of the Navier-Stokes coverage skipped, which is that OpenAI's release included a machine-checkable Lean 4 proof. Simon Willison covered the release the same day. That is the second headline result in a week to arrive with a formal artifact attached, after the Fermat formalization on September 4. In the narrow set of domains where output is checkable, a proof converts a claim about a model into a pass or a fail. Most domains are not in that set, which is the limit worth holding in view.
Shopify Engineering published its reversal back to separate Swift and Kotlin codebases, citing the cross-platform ceiling it kept hitting. Write once, run anywhere was the pitch in 1996 with Java, and the argument then rested on the same headcount math it rests on now: one team, one codebase, two platforms. The angle that did not make the writeup is that code generation has quietly cut the marginal cost of maintaining a second native app, which takes a chunk out of the strongest reason to accept a shared runtime's ceiling at all.
DeepSeek published V4.1-Flash weights, and the flash tier now arrives at 512GB. Nothing under a multi-GPU box or a large unified-memory Mac will hold it. The class names have stopped describing the hardware, which makes flash a statement about latency targets rather than about what it costs to serve.
Of everything this week, the one to carry is SoL-Pi: a harness component with a name, a version number and a vendor, shipped on its own schedule with no model attached.