Cloudflare released Clef this week, a family of open-weight decision models, alongside a reinforcement-learning platform for specializing them on a customer's own traffic. A decision model is not a chatbot: it takes a fixed set of options and returns a calibrated probability for each one in a single forward pass, which covers routing, moderation, tool selection, and retrieval gating. The category has been surfacing from independent builders for about a week. An infrastructure vendor shipping the weights and the training loop on the same day is a different kind of event, because it puts a floor under the pattern that a hobby project cannot.
Tuesday's edition argued that most agent loops contain a handful of steps that are judgments rather than writing, and that almost nobody keeps a record of what their system decided. Clef changes the shape of that problem rather than closing it. A calibrated probability per branch is a record by construction, and it arrives with a number attached, which is what an escalation threshold needs in order to be a threshold instead of a guess. The judgment is still a judgment. It is now a logged one.
The cheapest test of the same thesis came from an independent developer the same day. Version 1.2 of Jeff-Qwen3.5-0.8B adds nine LoRA adapters to a 0.8B decision model and reports decisions running 38 times faster with 8.7 points better accuracy when the small model sits in front of Qwen3.8-27B, at under two gigabytes of additional memory. Version 1.0's weakness was zero-shot generalization, and task-specific adapters are the fix the author landed on. The whole experiment runs on hardware that is already in the room.
Pi reached 1.0 this week with a companion release, Pi Durable, aimed at persistence across long-running agent runs. Pi has been winning local-harness comparisons for two weeks, and a stable version plus durability work plus third-party extensions is what harness consolidation looks like from the outside. One of those extensions lets a local Qwen 27B skip the reasoning phase and answer directly, which turns "should this request reason at all" into an explicit step rather than a sampler trick. Separately, an open-source router for coding agents called Weave Router claims that ensembling across models beats any single model and reaches Astra-level performance. Routing has been sold as a cost play for most of a year. This claim is about quality, and it belongs to the harness rather than to the application.
Matthew Green spent the week on the half of agent security that per-agent sandboxing does not reach. An agent worm has two parts: a payload that hijacks an agent, and an agent that will pass that payload along to the next one. A sandbox contains escape. It does nothing about propagation, because propagation travels on the message channel between agents, and most systems treat that channel as trusted internal traffic rather than as input of unknown origin. Simon Willison's writeup is where the argument surfaced.
Clef's open-weight release wasn't the only thing that moved on the "open weights" question this week. The de-alignment tool Heretic got a PewDiePie feature, a distribution event pulling in the other direction: expect a wave of abliterated re-uploads, and a Hugging Face "open weights" label that increasingly describes a starting point rather than the artifact a user ends up running.
Of everything this week, the one to carry is Green's split between escape and propagation. Clef will have competitors inside a quarter, and Pi's 1.0 will be 1.3 by December. The claim that an agent-to-agent message is untrusted input does not expire with the next release, and the multi-agent pipelines getting wired together right now are being wired on the assumption that it is false.