Bounded decisions should never enter the LLM loop

Agent cost work is usually framed as model compression. It is really two different jobs. On 18 September 2026, Cua open-sourced CUA-S1-FORMS: a 706,048-parameter, 2.8MB form decision model that fills an entire form inside a single forward pass of roughly 50 milliseconds. Handing the same form to a general LLM agent takes 23 rounds and 39.6 seconds. Our judgment: the second path to cheaper agents is not making the big model smaller, it is keeping bounded decisions out of the loop entirely. That path changes no model. It changes which decisions you put inside the loop.

1. The facts: a 706K model replaces 23 rounds of reasoning

Cua, founded by Francesco Bonacci, released CUA-S1-FORMS on 18 September as the first checkpoint of its System One model family: 706,048 parameters, a 2.8MB byte-level transformer, MIT licensed, with the model card published as cua-ai/cua-s1-forms on Hugging Face. Given a single GUI form field and a candidate value set, it decides fill, check, click or skip inside one forward pass, with no chain of thought, no multi-turn tool calls and no LLM API call. Sources: trycua/cua repository and the model card.

The comparison is the point. On the same form, a general LLM agent needs 23 rounds and 39.6 seconds, while the task-specific decision model needs one forward pass of about 50 milliseconds. The model card also reports 99.7 percent decision accuracy against a general LLM API baseline of about 83.6 percent. Source: AI Weekly coverage of the release.

Citation discipline: both accuracy figures come from the model card, so they are vendor-reported, and they cover one narrow task rather than general capability. What holds up is the mechanism and the order of magnitude. Bounded decisions can be handled by very small models, and can is a claim you verify on your own evaluation set.

2. Our judgment: two paths to lower cost, and the second one is ignored

Path A compresses the general model: quantization, distillation, caching, all aimed at a lower price per token. The ceiling on that path is set by the vendor pricing and release cadence. Enterprise buyers choose within it, they do not design it.

Path B moves bounded decisions out of the loop. It changes the architecture rather than the model: which judgments genuinely need general reasoning, and which ones one forward pass can settle. The return on that path is not set by vendor pricing. It is set by how well you decompose your own tasks.

Our judgment: path B gets ignored because it does not look like an AI project. It looks like process mapping. But its leverage is higher. Moving one judgment out of the loop removes rounds, context length and retries at the same time, and none of those three are visible on a per-token price sheet.

3. The test: which decisions to pull out

Three questions have to answer yes at the same time before a decision is a candidate:

The negative case is just as clear. Long-tail exception handling, judgments that require cross-system semantic understanding, and open-ended generation with unverifiable output belong in the general loop. Pulling decisions out does not make an agent smaller. It reserves the loop for the parts that actually need it.

4. Rerunning the economics: cost per completed task

One form-filling task inside a general agent is 23 rounds of inference: every round carries context, every round can retry, and every retry is billed again. Replaced by a single forward pass, marginal token cost goes to zero and latency drops three orders of magnitude. That is not a 10 percent optimization, it is a change in the cost structure.

The replacement is not free. You take on a small model, its evaluation set and its versioning, plus the risk that the bounded task definition drifts: interface changes, new candidate values and revised business rules all invalidate it and force retraining or revalidation. So the test is not whether it can be done, but whether the decision repeats often enough and is defined stably enough to be worth owning.

Our judgment: unit economics should be measured as cost per completed task, not as token price. Token price only lets you see the spread between model tiers. Cost per completed task lets you see structural space of a different kind, the space where a decision should not have entered the loop at all.

5. Deployment checklist and this month move

The smallest useful move this month: pull a week of agent production logs, plot the distribution of loop rounds, find the decision with the most rounds and test it against the three conditions above. Then answer one question: how many rounds in your agent loop are spent answering a question whose candidate set has ten options?

References: Cua open-source repository github.com/trycua/cua (HTTP 200) | Model card huggingface.co/cua-ai/cua-s1-forms (HTTP 200; source of the 706,048 parameters, 2.8MB size, MIT license, roughly 50ms forward pass, the 23-round and 39.6-second comparison, and the 99.7 versus 83.6 percent accuracy figures, all vendor-reported) | AI Weekly coverage aiweekly.co alert (HTTP 200).