Score the next step. Do not write a paragraph about it.
A closed choice should return a score you store with the work, and a weak score should stop the path instead of becoming a paragraph.
If the next step is a yes or no, stop asking a chat model to talk its way through it.
On October 1, 2026, Cloudflare released Clef and Clef-flash, two open-weight decision models under the Apache 2.0 license. The weights are on Hugging Face, and Cloudflare also hosts the models on Workers AI. You pass in a situation and a set of typed questions. The model returns a probability for an allowed answer. It does not write a paragraph. On Cloudflare's own latency table, the median is 209.3 ms for Clef and 38.8 ms for Clef-flash. (Cloudflare, October 1, 2026)
That is the spine of a running agent. Most failures that matter are not bad sentences. They are a guessed branch that the next step treats as fact. A closed choice needs a score you can store. A weak score should stop the path. Foundry Solutions builds that gate into the system: the choice, the score, and the stop live in the same memory as the work, and a person sees the case when the score is not strong enough to continue.
What does a decision model return?
A chat model returns language. A decision model returns a classification with a probability, over answers you already allowed. Cloudflare describes the pattern with a message that needs a yes or no on urgency, a choice among paths, and a severity score. The calling code can route, escalate, or defer. There is no paragraph for the next step to misread. (Cloudflare, October 1, 2026)
Clef and Clef-flash are built for that shape. They are open weight, so the gate can run on infrastructure you control, not only as a hosted call. Cloudflare's table also puts Clef-flash well under the latency of the decision model it compares against, which matters when the check sits on every step rather than once at the end. (Cloudflare, October 1, 2026)
The output you want is small on purpose. Which tool. Whether the case is weak. Whether a person has to see it before the agent goes on. If the answer cannot be written as an allowed choice, it is not this step. Send that work to a model that is supposed to write, or send it to a person.
Why do closed choices matter in a running agent?
An agent is a chain of branches. One of them picks a tool. One of them decides the case is clear enough to continue. One of them decides the work is finished. When those branches are hidden inside a chat reply, the next step has nothing solid to read. It reads tone. It treats a hedge as a yes. The error is quiet, and then it is finished output.
A score changes the contract. The branch is a field, not a sentence. Later steps can see the number and the allowed answer that won. So can a person reviewing the run. Companies of every size hit the same bug here: the demo sounds fluent, and the live path has no record of why it turned.
Foundry Solutions wants that record in shared memory, next to the work, not in a log nobody opens. The score is part of the state. If a later step cannot see it, the gate did not happen.
When should a weak score stop the work?
Set a line before you ship the agent. Above the line, the allowed answer stands and the agent continues. Below the line, the path stops. A person sees the case. Do not let the next step invent confidence the score did not give you.
Cloudflare is right that a typed answer can be used by code without a person on every easy case. That is the point of a fast gate. It is not a reason to let a weak score through. "The model was unsure, so it wrote a careful paragraph" is how unsure becomes action. Stop instead.
Two other launches this week sit next to this gate. They do not replace it. On October 7, 2026, Anthropic released Claude Haiku 5.5, a small model for repetitive, high-volume steps and for subagent work beside larger Claude models, at around 75 percent less cost to run than Haiku 4.5. Use that class of model on the narrow writing step after the branch is settled. Do not use it to talk the branch into existence. (Anthropic, October 7, 2026)
On October 6, 2026, Claude's default in Google Docs, Sheets, and Slides became "Ask before edits": each change is a card, and a person approves it before it lands. A second mode applies edits without stopping. (Anthropic, 9to5Google, October 6, 2026, 12:10pm PT) That stop is the same idea at the end of the path. A weak score stops a branch. A person stops a write to a real file. Generation, however cheap, sits in the middle.
What to do this week
- Pick one yes-or-no, or one which-path, inside an agent you already run. Write the allowed answers. Take that branch away from the chat model that currently explains it.
- Score the branch with a decision model such as Clef or Clef-flash. Store the winning answer and the score in the same record as the work, where the next step can read them. (Cloudflare, October 1, 2026)
- Choose the line that counts as weak. Below it, stop and hand the case to a person. If that same agent can change a real file, keep a person in front of the write. If you use Claude in Google Docs, Sheets, or Slides, leave "Ask before edits" on.
Foundry Solutions: AI that transforms business. Leading strategy. Building agents. Engineering infrastructure. Delivering capabilities.
If there is a job you want an agent to take on, start at /contact/.
If there is a job you want an agent to take on.
Foundry Solutions: AI that transforms business. Leading strategy. Building agents. Engineering infrastructure. Delivering capabilities.
Contact us →