A planner helps browser agents use practiced web skills
Back to Basics: A Skill-Centric Approach to Generalist Web Agents · arXiv:2610.35823
In plain words
Browser agents have to do many kinds of things on websites, and a task can fail if one-step goes wrong. This paper introduces MiniWeb, a set of websites and tasks that lets researchers check which interactions agents can complete.
The authors trained a small open model to practice individual interactions, then paired it with a planner that breaks bigger requests into steps. Practice alone improved scores on individual interactions but barely changed success on full tasks; the combined system did better on held-out tasks. The study tested one model and a limited set of websites.
The research
A browser task can depend on several different website interactions, so practicing one interaction may not be enough to complete a whole request. The question is whether an agent can turn that practice into better performance on tasks it has not seen during training.
MiniWebAgent pairs a planner that breaks web tasks into steps with an executor trained to carry out reusable interactions.
The authors first trained a single Qwen3.5-4B agent on examples of individual interactions. It improved on generated single-interaction tasks, but barely changed success on a separate set of human-authored held-out tasks. They then paired a trained executor with an untrained planner that breaks requests into steps.
On 62 held-out tasks, this combined agent passed 58.1%, versus 38.7% for the untrained single agent, averaging three runs.
This supports the idea that planning can help apply practiced interactions across longer tasks, though the comparison does not isolate the planner’s and executor’s contributions.
These results come from one compact open-weight model and a limited set of websites and tasks. MiniWebAgent combines an untrained planner with a trained executor, so the comparison does not show how much of the gain comes from either part alone.
What the evidence shows
- Make interactions measurable
MiniWeb is a benchmark for testing browser agents on website tasks. It contains 369 human-authored tasks built from 37 reusable interactions, or primitives, across seven skill families.
Each task marks which interactions it requires, so the benchmark can report not only whether a whole task passed but also which component interactions succeeded.
- Find a shared weak spot
That breakdown showed a consistent weakness among the ten untrained open-weight agents tested: drag and gesture was the weakest skill family for each. Across agents, 12.4% of those interaction instances passed, compared with 38.0% for the next-weakest family, out-of-page input and output. The pooled figures reveal a gap in this tested group, not a universal weakness in browser agents.
- Practice alone falls short
MiniWebGen creates short training examples for individual interactions and retains them when a backend check verifies the intended result. Training Qwen3.5-4B on these examples raised success on 103 generated single-interaction tasks from 55.3% to 80.3%.
On a separate set of 62 human-authored tasks from held-out sites, however, the trained single agent passed 39.2%, almost the same as the untrained model’s 38.7%.
- Add a planner to execute
MiniWebAgent is a browser agent that pairs an untrained planner, which splits a request into steps, with an executor trained on MiniWebGen examples. On the 62 held-out human-authored tasks, it passed 58.1%, compared with 38.7% for the untrained single agent; both figures average three runs.
The results show that the combined setup did better on these tasks, but do not attribute the full difference to either component alone.
- Compare training data
The researchers also compared the skill-trained executor with one trained on a similarly-sized set of complete-task trajectories, keeping the planner-executor setup and training recipe the same. The skill-trained version scored higher on all four evaluations, but only its Online-Mind2Web lead remained statistically significant after correction for multiple comparisons.
This supports skill-focused training in the tested setup, while leaving the strength of the advantage across other tests uncertain.
Deep
- Original paper authors
- Minh Pham-Dinh, Md Farhan Ishmam, Hernan Yupanqui, Reaz Hassan Joarder, Kenneth Marino
- Submitted to arXiv
- License
- CC BY-SA 4.0