Open Decision Models, Agent Swarms, and the Undo Button

James Turner

Welcome to 48 Hour AI. I'm James Turner, and it's Monday, October 5th. Let's start with a detail that's easy to skim past. If your app is built on TypeSafe's SDK and you want to swap in a different decision model, you change two strings. The endpoint, and the model name. That's the whole migration. That one fact explains a lot of what happened over the last 48 hours.

James Turner

Two open source families of decision models just landed, and both speak the same Jev API that TypeSafe uses. The first is Kev. It's four models, from 0.8 billion to 27 billion parameters. It's built as a drop in replacement for Jev. Code, weights, and evaluation reports are all public. The caveat is that the full training corpus for the 27B model and some of the evaluation data are still private. So it's open, but not open all the way down.

James Turner

The second is Clef, along with a smaller version called Clef flash. Both are on Hugging Face under an Apache 2.0 license, and you can run them locally. The release also comes with a new reinforcement learning fine tuning platform.

James Turner

Now, what is a decision model? It's not a chatbot. It doesn't write paragraphs. You hand it a situation and a set of options, and it classifies, returning a typed answer with probabilities attached. Kev does exactly that and nothing more. It scores the options you supply. It doesn't explain itself, and it won't go fetch a fact you forgot to include.

James Turner

That narrowness is the point. Those probabilities let an agent do three things in code. It can gather more context, it can take an action, or, if confidence is low, it can hand the decision to a human. That last move is a plain if statement, instead of a vibe.

James Turner

Here's my read, and it's interpretation, not something either release claims. Developers seem to be splitting agents into two parts. One is a small, fast steering wheel that picks the next move. The other is the big generative model you call only when you need real thinking or writing. If the steering wheel is a compact classifier you run yourself, you're not paying a giant model to decide whether to click the blue button or the green one.

James Turner

Last week we covered OpenAI's Decisions API. This is the open answer to that idea, and it shipped within days, in a format built so you can leave without rewriting your app. That's what a commoditizing layer looks like.

James Turner

Now, the steering wheel is one question. Who's driving is another. A piece circulating this weekend called The Dot and the Swarm argues that AI systems have outgrown the management structures humans designed for them. It points to Meta's Muse and OpenAI's Dots as agents that act autonomously and correct human errors without much input. It also says OpenAI used a swarm model to attempt the Navier Stokes problem, a grand open challenge in mathematics, by coordinating vast numbers of self directed agents. Notice the word attempt. The summary I have doesn't say it was solved.

James Turner

There's a researcher's view that fits here. Alex Zhang, an MIT researcher behind Recursive Language Models, was on the Latent Space podcast. His argument is that we leave a lot of capability on the table by wrapping powerful models in primitive harnesses. He says Claude Code, Codex, and most coding agents are more similar than they look. So the gains may come from better harness design, not just bigger models. And he asks what happens when you can throw 10,000 agents and tens of millions of dollars at a single problem.

James Turner

Then there's the open science side. Nathan Lambert and Tom Zick launched Trillium Labs, a nonprofit for the open science of frontier AI. The bet is that hard technical problems need more eyes, and that AI is getting too closed for the scientific community to do its job. They're building open post training recipes, and plan to grow into open infrastructure for studying recursive self improvement, reward hacking, and multi agent systems. They're hiring, fundraising, and looking for compute. Ai2 also shipped two open releases. AstaBrief 8B generates fast, cited scientific reports. Olmo core 3 is a redesigned framework for training large mixture of experts models.

James Turner

Put those together and you can see the pressure. Agents are getting more autonomy, swarms are getting bigger, and the tools to steer them are going open. So the next question is what happens when one of them gets it wrong.

James Turner

On October 2nd, we looked at NVIDIA's Sentry kill switch, hardware an agent can't reach. I asked what happens to half finished work when it pulls the plug. A CACM blog post titled AI Agents Need an Undo Button takes on that question.

James Turner

Its core idea is that as agents gain the ability to change real systems, we have to design not only for preventing harmful actions, but for recovering after they happen. Think about what a kill switch really does. It stops the agent. It doesn't tidy up. If an agent is halfway through changing a database or a cloud setup, you can be left with a system in a partial, inconsistent state. Stopping is not the same as safe.

James Turner

So the recovery toolkit looks like what database engineers have used for decades. Snapshot the state before the agent acts. Log every change as a transaction. And build deterministic rollback, so you can reverse what the agent did, whether it was a mistake or something malicious.

James Turner

And look at how this connects to today's other stories. A decision model that returns probabilities can tell an agent when to stop and ask a human. A kill switch can stop it cold. An undo button cleans up what's already done. Those are three layers of control, and only one of them is about intelligence.

James Turner

If your agents are going to run in swarms of thousands, rollback isn't a nice feature. It's the thing that decides whether you can afford to let them act. That's the show for today. I'm James Turner, and I'll see you in 48 hours.