Decido-1 (web-trained): choosing the next action in a browser agent
A browser agent makes one decision at every step: given the page, what should it do next, and to which element? Fast browser agents send this decision to a small decision model instead of a large LLM. We trained a version of Decido-1 for this decision, Decido-1 (web-trained), put it into Jev Ultrafast, Browser Use's open-source browser agent built around TypeSafe's Jev, and compared it with Jev on exactly the same requests. The released Decido-1 has no web training. On 1,500 steps from the Mind2Web test sets, Decido-1 (web-trained) chooses the right action on the right element 78.6% of the time, against 59.9% for Jev 1.14 and 54.4% for Jev 1.13. It is ahead on all three test sets, including websites and whole domains that were not in its training data. On live websites the two are on par today (74% vs 75% of 100 new tasks); we report that too, along with what the live Decido system adds.
1Offline results
Harness. We used Jev Ultrafast, Browser Use's open-source browser agent built around TypeSafe's Jev (commit 8251e27). At every step it turns the page into a table of the elements you can interact with and asks the decision model two questions: which operation (click, type text, select, scroll up or down, wait, done or blocked), and which element. A small LLM writes text only when the operation is "type text". For the offline test we ran the agent's own request builder, unmodified, on Mind2Web's saved pages and sent the identical requests to each model, Jev through the TypeSafe API.
The steps come from Mind2Web, a benchmark of real tasks recorded by people on real websites: 500 steps from each of its three test sets, with 50 candidate elements per step (the correct one and 49 others from the same page). A step counts as right only if both the operation and the element are right.
| Model | New tasksKNOWN WEBSITES | New websitesUNSEEN | New domainsUNSEEN | All 1,500 | Element | Operation |
|---|---|---|---|---|---|---|
| Decido-1 (web-trained) | 79.4 | 79.6 | 76.8 | 78.6 | 80.7 | 94.1 |
| Jev 1.14 | 62.4 | 59.0 | 58.2 | 59.9 | 62.3 | 86.8 |
| Jev 1.13 | 55.8 | 52.0 | 55.4 | 54.4 | 56.9 | 80.0 |
| Decido-1 (released, no web training) | 42.0 | 39.8 | 45.8 | 42.5 | 44.8 | 65.2 |
TABLE 1 Step accuracy (%) on 1,500 Mind2Web test steps, 500 per test set. Bold: best per column. "Unseen": no website from that test set was in Decido's training data. Element and Operation: each part right on its own. Failed requests count as wrong (Decido: 6 over its 32,768-token limit; Jev: 7). Jev scored through the TypeSafe API, November 2026.
FIG. 1 Step accuracy (%), all 1,500 steps. Dashed line: Jev 1.14. Same agent, same requests; only the decision model differs.
Head to head. On the same steps, Decido-1 (web-trained) is right where Jev 1.14 is wrong 356 times; the reverse happens 75 times.
| Test set | Steps | Decido-1 (web-trained) − Jev 1.1495% CI | Only Decido right | Only Jev right | Decido-1 (web-trained) − Jev 1.1395% CI |
|---|---|---|---|---|---|
| New tasks (seen websites) | 500 | +17.0 [+12.8, +21.2] | 110 | 25 | +23.6 [+19.2, +28.2] |
| New websites | 500 | +20.6 [+16.2, +25.0] | 125 | 22 | +27.6 [+22.8, +32.4] |
| New domains | 500 | +18.6 [+14.4, +23.2] | 121 | 28 | +21.4 [+16.8, +26.0] |
| All | 1,500 | +18.7 [+16.2, +21.3] | 356 | 75 | +24.2 [+21.7, +26.9] |
TABLE 2 Paired differences in step accuracy (points), bootstrap 95% confidence intervals over steps.
Where it still slips. Accuracy depends on how many elements the agent offers: 85.0% with 25 candidates, 77.3% with 50 and 74.7% with 100 (measured on 300 of the steps, the first 100 of each test set). If the elements are shuffled instead of listed in page order, accuracy drops to 71.0% and the chosen action changes on 20% of steps (not measured for Jev). Real pages arrive in page order, but this is the model's main weakness. Typical misses are near-ties, such as two filters with the same label.
Decido-1 (web-trained) is the same size as the released Decido-1: a 26B mixture-of-experts model that uses 4B parameters per token.
2Live example
Jev Ultrafast with Decido on the live Google Flights site, start to finish: enter the route and date, search (reloading once when Google shows an error page), switch to the cheapest results, sort by price, pick the cheapest flight and stop on its booking page. 24 actions in 26 seconds, real time and unedited.
VIDEO 1 A demo, not part of any evaluation. It uses the live Decido system from Table 3 plus one extra rule: wait while results are loading. Google showed new prices after the list was re-sorted (the cheapest went from $345 to $408); the agent picked the cheapest flight on the sorted list. We recorded four runs. By our reading of the recordings, the first two failed for both Decido and Jev 1.14 (a flight picked before the cheapest list had loaded, or Google's error page). We then added the wait rule to the Decido system only; the next two runs succeeded for both systems (Jev's ended on a $345 fare, at a different time).
3Live tasks
Offline replay checks one decision at a time with the earlier steps given. A live task also needs the agent to recover from its own mistakes and to know when to stop. We wrote 100 new tasks on live websites, mostly navigation on Wikipedia, GitHub and documentation sites. We froze the systems and the success check for each task (a rule on the final page's address or text) before the first run, and ran each system twice per task in Jev Ultrafast. For these runs the agent had three small changes to how it reads the page, the same for every system: it skips elements hidden under cookie banners, it isn't thrown off by unrelated page text changing, and it reports values already chosen in tag-style inputs. Text for "type text" steps came from Qwen3-4B for every system.
The Decido system is more than one model. Decido-1 (web-trained) picks each action; before stopping, the released Decido-1 is asked whether the goal is visibly done; and three fixed rules apply: ignore a stop when the action model is at least 90% sure of another action, wait once instead of repeating an action that left the page unchanged, and give up after the same action three times. Jev ran on its own, as the agent is designed.
| Decision system in Jev Ultrafast | Tasks completed100 TASKS × 2 RUNS |
|---|---|
| Decido (web-trained + released Decido-1 + rules) | 74.0 |
| Jev 1.14 | 75.0 |
| Jev 1.13 | 72.0 |
TABLE 3 Live websites, % of 200 runs that passed the task's check. Decido − Jev 1.14: −1.0 points (95% CI −9.0 to +7.0); Decido − Jev 1.13: +2.0 (−5.5 to +9.5).
Live, the two are on par: the 19-point offline lead does not carry over yet, even with the extra stop check. Most of Decido's live failures are stopping too early, when the page mentions the target before the task is done, and repeating an action that does nothing. We are training on the agent's own live runs to fix both and will report the result on a new set of tasks, whichever way it goes.
4Limitations
- Decido-1 (web-trained)'s training data is separate from the test sets: no shared tasks or near-duplicate goals, and no websites from the "new websites" and "new domains" sets; "new tasks" uses websites it was trained on, by design. Jev's training data is not public. No outputs from Jev or any closed model were used.
- Offline replay gives the model the true earlier steps and 50 candidate elements; a live agent sees its own history and every element on the page. The text typed into fields is not scored offline; live, it can affect whether a task passes, equally for every system.
- Decido-1 (web-trained) is not yet released; the released Decido-1 scores 42.5% on these steps.
5Work with us
If you're building a browser agent and want to try Decido-1 (web-trained) on your tasks, or want a decision model trained on your own agent's steps, write to us.