# rtrvr data: browser-agent training data, evals, and environments

> Canonical page: https://rtrvr.ai/data

rtrvr has served 35,000+ users across more than 7 million live-web runs. We turn those runs into training data, evals, and browser environments for agents that need to act, recover, and verify.

## Production proof

- **7M+** live-web runs
- **35K+** users served
- **81.4%** Halluminate Web Bench, ranked #1 ([source](https://rtrvr.ai/blog/web-bench-results))
- **3** complete runs to inspect

## What one run contains

- **Page state:** An enriched accessibility tree records every interactive element, on screen or off, with a stable ID.
- **Plans and tools:** The prompt, planner reasoning, generated code, API and network calls, and resolved arguments.
- **Failure and recovery:** Typed tool results preserve the exact error, the next attempt, and the path that finally worked.
- **Verified outcome:** Action, step, and task-level checks. Completion is settled against a fresh page state.

## Capabilities this data can measure

- **Find the right control:** Pages can contain thousands of interactive elements, including controls outside the viewport.
- **Keep the state:** Long tasks cross tabs, redirects, logins, apps, and human approval turns.
- **Recover in place:** A stale element or failed tool call should repair the plan, not end the run.
- **Know when it is done:** The agent must check the result instead of treating its own last action as proof.

## Production task mix

Measured over a recent one-week production window, deduplicated to distinct tasks and grouped by function.

- **Data extraction & scraping, 28.8%:** Pull structured records from listings, dashboards, and directories, usually into a sheet. Typical horizon: 10–100+ actions, often fanned across many tabs.
- **Job applications, 19%:** Applications through Ashby, Greenhouse, Lever, and Workday, including files, fields, and essay answers. Typical horizon: 10–40 actions on one long form.
- **Docs, Sheets & Drive ops, 12%:** Creating and editing docs and spreadsheets, and moving results between the browser and Google Workspace. Typical horizon: 5–50 actions, frequently as the sink of another task.
- **LinkedIn networking & outreach, 11.2%:** Finding people, verifying identity, sending connection requests and messages, and confirming delivery. Typical horizon: 5–20 actions per contact.
- **E-learning & assessments, 8.2%:** Course modules, quizzes, and certification flows on SkillsBuild, SCORM players, and publisher courseware. Typical horizon: Very long, often 50–500+ actions across many pages.
- **Business forms & ops, 7.5%:** Government portals, invoicing systems, CRMs, and internal tools, including checkout flows that pause for approval. Typical horizon: 10–60 actions, often multi-step wizards.
- **Content generation in-page, 6.9%:** Drafting and posting inside the target app's own editor, including CMS bodies and rich-text fields. Typical horizon: 5–25 actions.
- **Social media actions, 6.4%:** Posting, replying, reacting, and scheduling across social platforms, usually from a list of targets. Typical horizon: 3–15 actions per target, repeated.

## Complete sample runs

### Job application on Ashby

- Prompt: `apply to this job`
- Site: jobs.ashbyhq.com
- Model: deepseek-v4-flash
- 15 steps, 1 plan script(s), 5 sub-agent decisions
- 7 page states containing 722 interactive elements and 61 links
- What it shows: One planner pass writes the program; twelve browser calls execute without another model round-trip. Every field and file upload resolves to the exact interactive element in the captured page state. A fresh page check verifies that the application was submitted.
- Caveat: This is a clean run with no failed actions. Use the Amazon sample for recovery behavior.
- Raw trajectory: https://rtrvr.ai/trajectory-samples/job_application/workflow.json
- Bundle: https://rtrvr.ai/trajectory-samples/job_application.zip

### LinkedIn identity check and outreach

- Prompt: `go send jerry wu of halluminate a linkedin dm of i am about to send you the latest versions of the trajectories`
- Site: linkedin.com
- Model: deepseek-v4-flash
- 4 steps, 1 plan script(s), 5 sub-agent decisions
- 6 page states containing 9,149 interactive elements and 1,551 links
- What it shows: The model checks the search results and resolves the right identity before it sends anything. The selected control is one of 2,040 interactive elements on the profile page. The final page state shows the sent thread, so stopping is tied to visible evidence.
- Caveat: The agent summary says `incomplete` while the final review says `done`. The bundle keeps that disagreement.
- Raw trajectory: https://rtrvr.ai/trajectory-samples/linkedin_dm/workflow.json
- Bundle: https://rtrvr.ai/trajectory-samples/linkedin_dm.zip

### Amazon purchase with human approval

- Prompt: `go order me small splenda pack on amazon`
- Site: amazon.com
- Model: deepseek-v4-flash
- 9 steps, 3 plan script(s), 3 sub-agent decisions
- 5 page states containing 29,915 interactive elements and 4,165 links
- What it shows: The approval question and the human answer are stored inside the trajectory. Two steps fail, the agent repairs the run, and the order is still confirmed. One Amazon page contains 8,699 interactive elements and 1,137 links.
- Caveat: This bundle was redaction-processed: the delivery name, ZIP and order number appear as `[REDACTED_*]` placeholders inside the trees.
- Raw trajectory: https://rtrvr.ai/trajectory-samples/amazon_transaction/workflow.json
- Bundle: https://rtrvr.ai/trajectory-samples/amazon_transaction.zip

All samples: https://rtrvr.ai/trajectory-samples/rtrvr-trajectories-all.zip

Each bundle contains `workflow.json`, `trees/<id>.json`, `trees/INDEX.json`, `SUMMARY.json`, and a schema README.

## Work with rtrvr

### Learn from production runs

Use a focused cut of the corpus for post-training, imitation, or offline policy learning.

- Generalist or specialist by function, vertical, site, horizon, or failure type
- Page states, plans, tool calls, errors, recoveries, and outcomes
- New runs, redaction, and delivery format scoped with your team

### Train against the live web

Co-design browser environments for models that learn by acting, recovering, and checking results.

- Task families and starting-state setup drawn from observed work
- Browser, API, network, and code actions in the same environment
- Rewards and terminal checks anchored to the actual end state

### Measure the misses that matter

Build evals around the tasks your model is close to solving but still fails in production.

- Task taxonomies grounded in real browser work
- Short tasks and workflows that run for 30–60 minutes
- Verifiers for success, recovery, safety, and efficiency

### Browser work is code work.

Our planner writes code against the page, executes it, receives typed errors, repairs the program, and stops only after a final verification. The same loop matters for coding agents: long-horizon plans, tool execution, recovery, and checked completion.

## Research

- Aug 2025: [SOTA on Halluminate Web Bench](https://rtrvr.ai/blog/web-bench-results). 81.4% across 323 tasks without screenshots.
- Dec 2025: [DOM intelligence architecture](https://rtrvr.ai/blog/dom-intelligence-architecture). How we build a richer, smaller page representation.
- May 2026: [Replacing tool loops with eval()](https://rtrvr.ai/blog/eval-is-all-you-need-agent-harness). Why the planner writes a program instead of one click at a time.
- Jun 2026: [Code as plan](https://rtrvr.ai/blog/code-as-plan-deepseek-flash-text-only-browser-agent). A text-only agent that executes, repairs, and continues.
- Aug 2026: [The self-improving action layer](https://rtrvr.ai/blog/self-improving-action-layer-roadmap). How production runs become better site knowledge.
- Aug 2026: [Nothing but the cache](https://rtrvr.ai/blog/deepseek-prefix-caching-browser-agents). Byte-stable observations and a 90% lower token bill.

## Improve agents on real browser work.

Bring us one task your model misses. We will build the data, browser environment, eval, or verifier needed to improve it.

Contact: arjun@rtrvr.ai

## Citation

```bibtex
@misc{rtrvr2026trajectories,
  title        = {Web Agent Trajectories from Production Traffic},
  author       = {{rtrvr.ai}},
  year         = {2026},
  howpublished = {\url{https://rtrvr.ai/data}},
  note         = {Production browser-agent runs captured as enriched accessibility-tree
                  observations with typed actions and verified outcomes}
}
```