Back to Basics: A Skill-Centric Approach to Generalist Web Agents

1. Introduction

• The authors ask whether compact agents should spend limited training budgets on more complete tasks or on the reusable skills those tasks require.Their approach trains on short-horizon, skill-centric trajectories and composes learned skills at inference time.Source: §1 · pp. 1–2
• MINIWEB comprises 65 websites and 369 human-authored tasks built from 37 reusable interaction primitives with skill-level verification.Its primitive-first construction provides a controlled setting to measure and practice specific skills.Source: §1 · p. 2
• Across ten open-weight agents, drag & gesture is the weakest family for every model, averaging 12.4% success versus 38.0% for the next-weakest family.The authors suggest end-to-end trajectory scaling may favor frequent interactions while leaving rarer reusable skills underpracticed.Source: §1 · p. 2
• 19.4 points: MINIWEBAGENT improves success on held-out MINIWEB sites, with gains of 9.3 points on WebArena-Lite, 20.0 points on WebVoyager, and 16.1 points on Online-Mind2Web without target-benchmark training.The approach pairs a trained executor with a planner that decomposes tasks into practiced interactions.Source: §1 · p. 2
• Fine-tuning Qwen3.5-4B on 3.1K generated trajectories improves held-out single-primitive success by 24.9 points, with gains across all skill families.The pipeline generates verified, skill-specific trajectories; its short-horizon training does not automatically transfer to full tasks, motivating a planner-executor agent.Source: §1 · p. 2

2. Related Work

• Web benchmarks range from small synthetic widget tasks to realistic self-hosted sites and broader but less stable live-web settings; many score whole tasks without attributing failure to a specific skill.MINIWEB embeds interaction skills in multi-page websites with persistent backend interactions, cross-site tasks, and per-primitive verdicts.Source: §2 · p. 2
• Figure 2 depicts MINIWEB's 65 websites by domain, held-out status, hubs, cross-site links, and data source.This site network supports the paper's contrast between isolated component testing and interaction skills embedded in connected websites.Source: §2 · p. 2, Figure 2 · p. 3
• Existing web-agent training commonly uses end-to-end trajectories, while MINIWEBGEN targets predefined primitives, including interactions that such data rarely contains.MINIWEBAGENT uses a planner-executor separation; the paper's contribution is training the executor through targeted practice and testing whether a planner composes those improvements.Source: §2 · p. 3

3. Methodology

• Current benchmarks do not expose component skills for separate evaluation, motivating tasks with explicit skills and per-skill verdicts.The methodology asks whether agents can reliably perform required skills and identify which one failed.Source: §3 · p. 3

3.1. MINIWEB

• MINIWEB is built bottom-up: recurring interactions become 37 primitives in seven families, which shape websites and tasks that chain them.The figure depicts tasks mined into primitives, sites built around those primitives, and tasks composed from them with a verdict per primitive.Source: §3.1 · pp. 3–4, Figure 3 · p. 4
• The environment spans 65 websites across 15 domains, with linked sites supporting cross-site workflows and both public-dataset-seeded and deterministic synthetic content.Half of the domains are absent from WebArena, Mind2Web, WebVoyager, and WebShop.Source: §3.1 · p. 4
• Annotators composed 369 tasks from supported primitives, recorded skill-segmented demonstrations, and refined the final set to 37 primitives across seven families.Tasks include dependencies and realistic instructions; six annotators produced them, including 43 cross-site tasks, and each task, verifier, and demonstration was manually reviewed.Source: §3.1 · p. 4
• MINIWEB uses parameterized verifiers for observable interactions and a VLM judge for semantic or reasoning outcomes, with task success requiring every required primitive to pass.On 549 validation cases, verifiers achieved 98.9% accuracy and the judge 98.2%; on agent episodes, their task-outcome agreement was 93.1%.Source: §3.1 · p. 4

3.2. MINIWEBGEN

• MINIWEBGEN summarizes demonstrations into site-agnostic primitive guides, samples controls and states, generates instructions, executes browser actions, and gates trajectories before keeping them.Its generated set contains 3,128 trajectories; the gate requires that attempts do not pass without acting.Source: §3.2 · p. 5

3.3. MINIWEBAGENT

• MINIWEBAGENT uses an untrained planner to delegate interaction steps to a trained executor, addressing the short horizon of single-primitive training trajectories.The planner observes the page but cannot act; the executor runs for at most 15 actions, while reasoning steps use the base model.Source: §3.3 · pp. 5–6
• Training replays generated trajectories through the browser-use harness and fits a shared LoRA adapter on the resulting action targets; held-out-site and external-benchmark trajectories are excluded.The training set includes 629 mistake-and-correction variants and 857 two-trajectory chains, totaling 27.6K steps with the plain trajectories.Source: §3.3 · p. 6

4.1. Drag & gesture is the weakest family in all web agent models.

• Drag & gesture is the weakest family for every untrained agent, averaging 12.4% success versus 38.0% for out-of-page I/O.In 72% of failed drag instances, agents reach the correct slider or canvas but fail during the gesture or immediately afterward; the bottleneck is executing and verifying the gesture.Source: §4.1 · pp. 6–7
• Agents have distinct skill profiles despite similar overall task success; task success correlates strongly with discrete selection, form transactions, and reasoning, but weakly with text entry.Aggregation is consistently difficult: counting is harder than reading a value for every agent, and finding an extremum is harder for eight of ten.Source: §4.1 · p. 7
• Across ten agents, success on n-primitive tasks stays within 3.8 points of p^n at every length, with mean one-primitive success p = 62%.The paper notes that every primitive in a chain must succeed, so primitive-level gains compound across longer tasks.Source: §4.1 · p. 7

4.3. Composing trained skills raises task success on unseen websites and benchmarks.

• 19.4 points on held-out MINIWEB, 9.3 on WebArena-Lite, 20.0 on WebVoyager, and 16.1 on Online-Mind2Web: MINIWEBAGENT improves task success without target-benchmark training.On held-out MINIWEB tasks, primitive success rises from 49.8% to 63.7%; on tasks with three or more primitives, success is 33.3% versus 14.5% for the untrained single agent.Source: Table 3 · p. 7, §4.3 · p. 8

4.2. MINIWEB practice improves single-primitive task generation, but not task success.

• On 103 generated single-primitive tasks from held-out sites, Qwen3.5-4B success rises from 55.3% to 80.3%, but task success on 62 human-authored tasks changes only from 38.7% to 39.2%.The trained single agent practices mostly one-primitive episodes with skill-naming instructions, unlike real tasks that chain primitives under one request; MINIWEBAGENT uses a planner to decompose those tasks.Source: §4.2 · p. 7

4.4. At matched size, our skill data matches or beats task-centric data.

• With a size-matched training corpus, the skill-trained executor outperforms the task-centric executor across all four benchmarks, with a significant +10.7-point gap on Online-Mind2Web and a +10.2-point gap on held-out MINIWEB (Holm p = 0.081).The task-centric executor does not significantly improve over the untrained planner on any of the four benchmarks, while the skill-trained executor does.Source: §4.4 · p. 8

• MINIWEBAGENT improves over the base model on held-out sites, WebArena-Lite, WebVoyager, and Online-Mind2Web without training on target-benchmark trajectories.Its trained executor adds to what planning alone achieves, most on held-out sites and Online-Mind2Web.Source: §5 · p. 8
• MINIWEB profiles 37 interaction primitives across 65 locally hosted websites, enabling per-primitive evaluation and skill-segmented data generation.Its tasks have skill-level verification.Source: §5 · p. 8
• Drag & gesture is the weakest skill family for every evaluated agent, while fine-tuning on generated single-primitive trajectories improves primitive execution but does not by itself raise multi-step task success.The model stops after one primitive, and removing that stop alone does not help.Source: §5 · p. 8
• MINIWEBGEN expands limited seed demonstrations into 3.1K verified trajectories for targeted skill practice.The paper concludes that synthetic-site skill practice combined with explicit separation of planning and execution helps compact open-weight agents on unfamiliar websites.Source: §5 · p. 8

• MINIWEB contains fewer human-authored tasks than Mind2Web and WebArena, and creating diverse tasks still requires substantial effort and creativity.MINIWEBGEN alleviates training-data scarcity but does not replace broader human-authored evaluation.Source: §6 · p. 8
• Closed-source frontier agents remain unevaluated because running them is expensive, leaving the study focused on small open-weight models.Source: §6 · p. 8
• The study does not fully resolve drag-and-gesture deficiencies and leaves specialized methods for mastering them to future work.Source: §6 · p. 8
• Scaling experiments were omitted because of compute and evaluation costs; future work should test larger datasets, greater website diversity, and larger models.Source: §6 · pp. 8–9

• The supplement covers the benchmark in Supplements A–E, agents and training in F–G, evaluation in H–I, additional results in J, and prompts in K.Source: Supplementary Material · p. 13

A. Mining candidate macros from web-agent benchmarks

• The section mines candidate macros from tasks in four web-agent benchmarks and uses the resulting candidates to build the final primitive set.It also reports the mining findings and explains how candidates became 37 primitives.Source: A. Mining candidate macros from… · p. 14

A.1. What is a macro?

• A macro is a site-agnostic, short-horizon unit for one recurring intent, structured as a verb_modifier pair between atomic actions and tasks.It reaches one verifiable terminal state and cannot be decomposed into smaller macros.Source: §A.1 · p. 14

A.2. Macro bank creation

• The macro bank is induced from four benchmark task sets, deduplicated with human adjudication, reassigned against a fixed vocabulary, and extended to MiniWeb domains.Additional macros are brainstormed under LLM assistance and human review to cover intents missing from the source datasets.Source: §A.2 · pp. 14–15

A.3. Mining statistics

• 52,207 macro labels cover 15,727 tasks across four benchmarks, using 136 candidate macros; tasks require 2.2–3.5 macros on average, and 22 candidates occur in only one benchmark.The labeled tasks include all WebArena and WebVoyager tasks, 12,251 WebShop instructions, and 2,021 Mind2Web tasks.Source: §A.3 · p. 15, Table 4 · p. 15
• WebShop supplies 78% of tasks and dominates pooled search and selection labels; without it, searching, filtering, and extracting information account for 61% of labels, while no label names dragging or drawing.Only 157 labels are price filters and 28 labels write a file; the price filters can be set by a slider.Source: §A.3 · p. 15, Table 5 · p. 15

A.4. From candidates to primitives

• The authors merged equivalent candidates, removed site-specific ones, separated reasoning from physical interaction, and added rare interactions, producing 46 initial primitives across 7 skill families and retaining 37 used at least three times.Added interactions include dragging a slider, drawing a signature, uploading a file, and editing a spreadsheet cell.Source: §A.4 · p. 16
• 131 of 136 candidate-to-primitive mappings were retained (96.3%, κ = 0.96; 99.7% of labels), yet drag-and-gesture accounts for 0.3% of mined labels versus 7.6% of MINIWEBGEN training-episode instances.Dragging reaches at most 2.1% of labels in Mind2Web; excluding reasoning, benchmark labels span effective counts of 3.4–9.6 primitives, compared with 32.8 for MINIWEBGEN trajectories.Source: §A.4 · p. 16

B. Interaction primitives and reasoning operations

• Each primitive instance pairs a physical interaction with an optional reasoning operation, while a single registry specifies its description, demonstration span, and VLM-judge rubric.The seven skill families inform a family-specialist variant that trains one specialist per group of families.Source: B. Interaction primitives and… · p. 16

C. Websites

• MiniWeb comprises 65 interconnected websites across 15 domains in a shared fictional world, with 21 sites seeded from public datasets, 44 using synthetic data, and 13 held out for evaluation.Six hubs carry actions across sites, including WebMail for receipts and notifications and SecureBank for payments and trades.Source: C. Websites · p. 18
• The catalog spans media, editing, and productivity sites, including music and video services, document and whiteboard editors, calendars, and cloud consoles.Entries describe user workflows such as playing and sharing media, editing documents, and managing cloud resources.Source: C. Websites · pp. 23–24
• Other entries cover business, civic, navigation, education, and utility workflows, with examples including project tracking, transit planning, course submissions, government services, and password management.Site descriptions identify each site's domain, train or held-out split, data source, task count, primitive count, and implementation.Source: C. Websites · pp. 24–26

D. Annotation interface

• Annotators explored sampled sites, mapped primitive instances and their order and data flow, wrote instructions, recorded tasks, and marked action spans and expected answers.The dashboard prioritized primitives with fewer tasks; optional LLM suggestions were not binding.Source: D. Annotation interface · p. 26
• Verifiers pinned instruction-dictated values, tested gold recordings and negative controls, while reviewers checked consistency, verification status, and whether tasks needed re-recording or deletion.The negative controls were an empty trajectory and an answer given without any action.Source: D. Annotation interface · p. 26

E. Grader validation

• Human demonstrations were retained only if current and accepted by the task verifier, leaving 183 of 369 recordings for grader validation.The excluded recordings included 67 tagged stale and 119 rejected by the current verifier.Source: E. Grader validation · p. 26
• Across 549 positive and negative cases, parameterized verifiers achieved 98.9% accuracy and the VLM judge 98.2%.The verifiers passed every positive; the judge passed 97.8% of positives, and each grader accepted 6 of 182 near misses.Source: E. Grader validation · pp. 26, 29
• On 3,689 untrained-agent episodes, the graders agreed on 93.1% of task outcomes (Cohen’s κ = 0.85).In a human audit of 120 episodes, the judge matched human verdicts on 92.2% and the verifiers on 88.9%; the audit prompted verifier fixes and removal of seven tasks.Source: E. Grader validation · p. 29

F. Observation and action spaces

• Each browser session starts from shared initial state but writes to its own overlay, with request logs and resulting state available to verifiers.The local Chromium setup simulates file controls, clipboard access, and two-factor codes or confirmations delivered through WebMail.Source: F. Observation and action spaces · p. 30
• The executor uses DOM-assisted target discovery under sandbox restrictions, while evaluated agents act from screenshots and indexed elements using one action per model turn.The shared harness adds drag, draw, and macro_done actions, removes page-bypass tools, and limits executor delegations to 15 actions.Source: F. Observation and action spaces · p. 30

G. Other architecture choices: family specialists

• Each family specialist starts from the pooled executor and learns its own family’s rows while staying close to the pooled executor on other groups’ rows.The steps are labeled by family group to form each specialist’s training data.Source: G. Other architecture choices: family… · p. 31
• The specialist objective combines cross-entropy on its own group’s data with a KL penalty on other groups’ data, weighted by β = 0.1.The penalty is non-negative and vanishes wherever the specialist and pooled adapters agree.Source: G. Other architecture choices: family… · p. 31
• Reasoning is not synthetically generated; its adapter uses the base model’s successful attempts on human reasoning tasks from training sites.The attempts are sampled four times per task at temperature 0.7.Source: G. Other architecture choices: family… · p. 31

H. Live-web evaluation sets

• The family-specialist agent scored 60.3% on all 643 WebVoyager tasks and 69.2% on the normalized 536-task subset.The subset removes impossible tasks, a site that blocked the network, and outdated task text; 17 of the 107 removed tasks were accepted by the judge.Source: H. Live-web evaluation sets · p. 31
• On Online-Mind2Web, the family-specialist agent scored 25.7% on all 300 tasks and 26.1% on the normalized 280-task subset.Nine sites were removed when at least half their episodes reported a network gate; only whole sites were excluded.Source: H. Live-web evaluation sets · p. 31, Table 12 · p. 32
• The live-web results are not comparable with results reported on the full benchmark sets because the evaluation uses normalized subsets.The subsets are fixed across configurations, and removals depend on websites or task text, not whether a configuration solved a task.Source: H. Live-web evaluation sets · p. 31

I. Compute and cost

• Training all executors took about 74 GPU-hours, while API calls behind reported results cost about $1.0K.The API total excludes collecting task-centric trajectories.Source: §I · p. 32

J.1. All configurations

• Family specialists trail the pooled executor on held-out tasks, WebArena-Lite, and Online-Mind2Web, but lead on WebVoyager without a significant difference.The appendix also evaluates a same-size web-trained model, Fara1.5-4B.Source: §J.1 · p. 32

J.2. Statistical tests

• On WebArena-Lite, the first run alone gives the trained executor a 10.4-point advantage over the untrained planner.The reported tests use Holm correction within each comparison question; figure stars use uncorrected p-values.Source: §J.2 · p. 32, Table 14 · p. 33

J.3. Why practice alone does not raise task success

• The trained single agent’s median episode shrinks from 18 to 9 steps on held-out tasks and from 40 to 8 on WebArena-Lite.It also shrinks from 80 to 9 on WebVoyager and from 100 to 14 on Online-Mind2Web; the untrained agent reaches its action budget in 56% of WebArena-Lite episodes.Source: §J.3 · p. 33

J.4. Stop-early control

• 58.1% on held-out tasks and 36.4% on WebArena-Lite are MINIWEBAGENT success rates, versus 39.2% and 20.3% for the trained single agent.Removing trained done steps yields 34.4% on held-out tasks and 22.5% on WebArena-Lite, still far below MINIWEBAGENT.Source: §J.4 · p. 33
• Without trained done steps, the single agent reaches its 40-action budget in 96% of held-out-task and WebArena-Lite episodes without significant success gains.On held-out tasks, success is 34.4% versus 39.2% with done steps (p = 0.42); on WebArena-Lite, it is 22.5% versus 20.3% (p = 0.44).Source: §J.4 · p. 33

J.5. WebArena-Lite by site

• Excluding Reddit and GitLab tasks, MINIWEBAGENT still gains 8.9 points over the untrained single agent on WebArena-Lite (p = 0.001).Its gain over the untrained planner is 4.8 points (p = 0.10).Source: §J.5 · p. 33

J.6. Model calls

• MINIWEBAGENT’s planner reaches its turn cap in 18%, 40%, 22%, and 30% of episodes across the four benchmarks.Planner turns do not count toward the site-action budget, so MINIWEBAGENT makes more model calls than the single agent.Source: §J.6 · p. 33

J.7. Robustness of the skill profiles

• 11.2% of drag instances pass the task verifiers, versus 34.8% for the next-weakest family.Under the judge, drag & gesture remains the weakest family for all ten agents even when only instances with previously passed primitives are counted.Source: §J.7 · p. 33
• Average success across agents stays within 3.8 points of the independent-primitive prediction p^n at every chain length.The comparison counts primitive instances only when all earlier primitives in the task passed.Source: §J.7 · p. 33
• Five of six agents with their own loops include a drag action yet pass at most 17% of drag instances.Fara-7B has no drag action, so its 0% drag success is by construction.Source: §J.7 · p. 33

J.8. Training and data details

• The skill executor trained for one epoch on 27,646 steps, while the task-centric executor used the same recipe on 27,644 steps from 3,795 trajectories.The skill data comprised 3,128 plain trajectories, 857 two-trajectory chains, and 629 trajectories with an injected and corrected mistake; the task-centric data came from 52 training sites. Its judge accepted 4,308 of 7,267 episodes and agreed with task verifiers on 87.1% of 310 human-task episodes (77.5% precision, 92.2% recall).Source: §J.8 · p. 35
• 6,384 candidate tasks yielded 3,128 training trajectories after dry runs, instruction writing, execution, deduplication, and review.Of 5,785 execution attempts, 3,653 succeeded on the first attempt; 4,011 executions showed the intended backend change, and 3,428 trajectories remained after deduplication and per-site caps.Source: §J.8 · p. 35

K. Prompts

• MINIWEBGEN stages and its macro judge use gemini-3.5-flash, while the MINIWEBAGENT planner and executor use Qwen3.5-4B.The section presents prompt templates with runtime values shown as placeholders and formats one-line JSON messages as indented text.Source: K. Prompts · p. 35

K.1. Macro induction

• Macro induction extracts intent-level actions independently against a frozen vocabulary, merges proposals through provenance-blind human review, then relabels all tasks against the same frozen vocabulary.Reported coverage and discovery curves are made order-free by averaging over random task or dataset orderings.Source: §K.1 · pp. 36–38

K.2. MINIWEBGEN: guide summarization

• Before task generation, MINIWEBGEN distills human demonstrations into site-agnostic primitive guides with concrete steps, variants, pitfalls, and completion signals.When no training-site demonstration exists, the guide is written from the primitive description, example, and rubric instead.Source: §K.2 · p. 38

K.3. MINIWEBGEN: instruction suggestion

• In stage 3, MINIWEBGEN’s sampler fixes the task parameters, and gemini-3.5-flash only phrases them as a natural-language instruction.The model runs in JSON mode at temperature 0.9.Source: §K.3 · p. 39
• Instruction suggestion supplies task parameters and phrasing styles to generate a full instruction and one imperative subtask per step, then checks option coverage and retries rejected outputs up to three calls.The checker rejects missing option words, answer-revealing content, field names or markup, and non-JSON or overly long instructions.Source: §K.3 · pp. 39–40

K.4. MINIWEBGEN: executor

• The executor asks a coding LLM to write a Python script for each step using a restricted action wrapper, live-page context, primitive guidance, and backend checks, with up to three attempts.Its range-slider guidance sets the value by clicking the track and requires clicking an apply button when the page does not apply the filter automatically.Source: §K.4 · pp. 40–42

K.5. Macro judge

• The macro judge compares the agent’s requested target with the reference, checks for unrequested changes, and prioritizes server requests over page state and agent claims.When screenshots are attached, visual criteria use them, but an accepted server response confirming state takes precedence.Source: §K.5 · pp. 43–44
• The macro judge requires every named element to be completed and counts an action only when the site accepts it, using screenshots when requests are silent or ambiguous.A 4xx/5xx response or an error form does not count as acceptance.Source: §K.5 · p. 44

K.6. MINIWEBAGENT: planner

• The planner delegates one page-grounded step at a time to a skill-family specialist, checks the resulting page, and avoids repeating failed or unchanged-page steps.It has a 40-site-action shared budget and at most 25 planner turns; specialists receive the planner’s instruction rather than the overall task.Source: §K.6 · pp. 44–45

K.7. MINIWEBAGENT: executor

• Each executor is a fresh browser-use agent on the same browser, limited to its assigned step and equipped with drag, draw, and macro-completion actions.In v6 it does not see the overall task and can take at most 15 steps, or fewer when the shared budget has less remaining.Source: §K.7 · p. 46