Add to ChromeThe page state keeps every interactive element, on screen or off, addressable with a stable ID.
Addressable stateTrain agents onknowledge work.
People use rtrvr for consequential knowledge work across live sites: research, applications, operations, documents, and outreach. We turn 7M+ production runs into trajectories, environments, evals, and verifiers for teams building agents that can finish the work—not just click through it.
Plain-text version for agents and crawlers: /data.mdTask
“complete this job application”
Page state105 interactive elements
01textboxName
id=6102textboxEmail
id=6603buttonUpload file
id=7504textboxEvaluation system
id=9705buttonSubmit application
id=112Action
Successupload_file · element 75Final check
Completefresh page stateOne exact page state, action, and final verifier from the sample below.
- 7M+
- live-web runs
- 35K+
- users served
- 81.4%
- Halluminate Web Bench, ranked #1
- 3
- complete runs to inspect
Production learning
Every run can improve the next.
The trajectory records what happened and proves the job finished. A success becomes a skill; a miss becomes a test.
branching state + action space → one verified trajectory01Smart DOM
02Code plan
03Act + recover
04Verified outcome
success → skill
failure → test
human rescue → repair
eval → better agent
Real browser work branches. Each trajectory preserves one exact path, then turns the outcome into the next skill, test, repair, or eval.
The instruction, planner reasoning, generated program, API calls, and resolved arguments stay together.
Reconstructable decisionTyped results preserve errors, retries, human rescue, and the path that finally worked.
A failure worth learning fromFresh page-state checks show whether the result—not merely the last action—was complete.
Proof of completionWhy this data matters
Where agents still fail.
The browser is a non-deterministic work environment: controls move, state crosses tabs and sign-ins, and a successful click can still leave the job unfinished. Production traces reveal the exact miss and whether recovery actually worked.
01Target
Perceive the right control
Pages can contain thousands of interactive elements, including controls outside the viewport.
02Continuity
Carry state across systems
Long tasks cross tabs, redirects, logins, apps, and human approval turns.
03Recovery
Recover without restarting
A stale element or failed tool call should repair the plan, not end the run.
04Proof
Verify the finished job
The agent must check the result instead of treating its own last action as proof.
Three complete trajectories
See how the work finished.
Follow the state, plan, tool arguments, errors, repairs, and final verifier. The viewer reads the same JSON that ships in the bundle.
Job applicationsjobs.ashbyhq.comdeepseek-v4-flash
Instruction “apply to this job”
Loading the run
A week of real work
Where agents earn trust.
A one-week slice of distinct production tasks. It shows where demand is real—and which capabilities need long horizons, recovery, or specialist data.
01Largest production slice28.8%
Structured research & extraction
Pull structured records from listings, dashboards, and directories, usually into a sheet.
0219.0%
Professional applications
Applications through Ashby, Greenhouse, Lever, and Workday, including files, fields, and essay answers.
0312.0%
Document & workspace operations
Creating and editing docs and spreadsheets, and moving results between the browser and Google Workspace.
0411.2%
Prospecting & professional outreach
Finding people, verifying identity, sending connection requests and messages, and confirming delivery.
058.2%
E-learning & assessments
Course modules, quizzes, and certification flows on SkillsBuild, SCORM players, and publisher courseware.
067.5%
Business systems & operations
Government portals, invoicing systems, CRMs, and internal tools, including checkout flows that pause for approval.
076.9%
Publishing in product editors
Drafting and posting inside the target app's own editor, including CMS bodies and rich-text fields.
086.4%
Social publishing & engagement
Posting, replying, reacting, and scheduling across social platforms, usually from a list of targets.
For labs and agent teams
Start with the corpus.Build what is missing.
Synthetic tasks are useful for scale. Production runs reveal the instructions, brittle states, human rescues, and end conditions that only appear when people entrust an agent with real work.
01 · Corpus
Train on consequential work
Curate production trajectories around the capability, industry, horizon, or failure mode your model needs next.
- Full page states, plans, tool calls, errors, recoveries, human turns, and outcomes
- Cuts for post-training, imitation, offline RL, or targeted error analysis
- New runs, redaction, and delivery format scoped with your research team
02 · Environment
Turn work into an environment
Recreate the starting state, action space, and terminal condition for online rollouts against real browser work.
- Task families and state setup drawn from observed production work
- Browser, API, network, and code actions inside one training environment
- Rewards and terminal checks anchored to the actual end state
03 · Evaluation
Measure the misses that matter
Build evals around valuable tasks your model is close to solving but still cannot finish reliably.
- Task taxonomies grounded in real knowledge work
- Short actions through workflows that run for 30–60 minutes
- Verifiers for completion, recovery, safety, and efficiency
Open research
How we made runs learnable.
The architecture behind page understanding, code-as-plan, recovery, verification, and cost. Each system makes a trajectory easier to reconstruct and learn from.
Three sample trajectories
Take apart three runs.
Each ZIP contains the full trajectory, captured page states, an index, summary counts, and the schema README.
Bundle contents
workflow.json full trajectory trees/<id>.json captured page states trees/INDEX.json IDs, URLs, and counts SUMMARY.json coverage and totals
Citation
@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}
}Work with rtrvr
Make one hard task reliable.
Bring one real task your agent cannot finish. We will scope the corpus, environment, eval, or verifier that moves it forward.
Scope the next capability