DeepSeek made V4 Pro 0813 official this week, and the weights are the less interesting half of the release. Alongside the launch post and the Hugging Face repo, the company published its own agent harness, the scaffolding it uses to drive the model through tool calls: prompt format, tool-call loop, system-prompt conventions. Builders have been inferring all three from traces. Monday's editorial made the point that an outside team reproducing DeepSeek's headline agent score on a public rig counted as unusual, and that the unusualness was the useful information. What changed by Friday is that the rig is published. Harnesses from this family have set the tool-call conventions that downstream frameworks copy, so the document travels further than the checkpoint does.
Qwen's first 3.8 model puts reasoning effort under prompt-level control, with low, medium and xhigh selectable per request instead of per deployment. That moves the depth-versus-latency call out of infrastructure, where it gets made once by whoever owns the deployment, and into the request, where it gets made constantly by whoever wrote the calling code. The other half of the release reads less well. The working Jinja chat template covering 3.5, 3.6 and 3.8 is community-maintained, not vendor-shipped, and a wrong chat template degrades output quietly rather than failing loudly. Some share of the model comparisons published this month is running on a template a volunteer fixed, which is how a bad benchmark gets into circulation without anybody noticing.
Geoffrey Litt put a name to the constraint sitting underneath all of this: understanding is the new bottleneck. Once generation is cheap, what a codebase runs short of is comprehension, and a team optimizing for output speed accumulates code nobody on it can reason about. The throughput limit lands with the people who have to hold a change in their heads long enough to approve it, and that is a smaller group than the one that can now produce changes. Litt's preference is legibility over clever compression, which is an old taste with a new reason behind it.
NVIDIA released Switchyard quietly alongside Nemotron 3.5 and it surfaced on its own this week: a Rust proxy that routes requests across providers and translates between the OpenAI and Anthropic request formats. Format translation was glue code every team wrote and maintained privately. A chip vendor shipping it as a library says the multi-provider setup is now the assumed shape, and that the translation problem is settled enough to become somebody else's dependency. It arrived in the same week as MCP-stama, a zero-dependency MCP server that ships as a single Rust binary, aimed at containers and CI runners where dragging a Node or Python runtime along is the objection. Two unrelated projects, one instinct: get the runtime out of the request path.
The least glamorous item of the week is an open systemd issue reporting that one log line can produce more than 49KB of journald writes on ext4 and more than 110KB on btrfs. That was tolerable when logs came from services that spoke occasionally. An agent loop narrating itself through a long task speaks constantly, and the amplification arrives as gigabytes a day on a VPS that was never sized for a tenant this chatty. Nobody chose that cost. It came with a default set for different neighbors.
Of everything this week, the one to carry is the harness. DeepSeek published the thing that says how to drive the model, and the frameworks that copy its tool-call conventions over the next few weeks will be copying a document instead of a guess.