Agentic trajectories → World models
Agentic trajectories for World models
World models needs long, continuous, real-world video with motion and depth cues. Here is why Agentic trajectories is the right raw material for it, what buyers typically spec, and how the rights are handled.
Why Agentic trajectories for World models
World models for agents learn environment dynamics: given state and action, what happens next. Agentic trajectories are the native training data because they contain the action channel that plain video lacks — the click, keystroke, or command that caused the next observation. For digital environments this yields state-action-state tuples at scale: UI worlds where a model learns that submitting a form navigates, deleting removes, scrolling reveals. The requirements are temporal and causal. Episodes must be long enough to contain multi-step structure; alignment between action log and observation must be tight enough that causality is unambiguous; and the environment must be versioned, because a world model trained across silently differing app versions learns a blurred, inconsistent world. Failures and dead ends are still valuable — the model should know what nothing-happened looks like — and coverage should span the state space, not just golden paths.
What buyers typically spec
Industry-typical ranges — a brief can and should deviate where the task demands it.
| Typical volume | Tens of thousands of episodes; long-horizon episodes weighted |
|---|---|
| Capture | Screen state (video or DOM snapshots) with aligned timestamped event log |
| Labels | Goal, per-step action semantics, terminal outcome; environment version per episode |
| Content | State-space coverage including dead ends and error states, not only success paths |
| Formats | JSONL events, MP4/screenshot observations |
A sample brief
The shape of a workable request — swap in your own numbers and conditions:
- Modality: full-resolution screen recordings with aligned event logs across web workflows.
- Volume: 20,000 episodes over 30 applications, versions pinned and recorded.
- Labels: goals, action semantics, outcomes; error states deliberately included.
- Rights: demonstrator consent; sandboxed accounts so no third-party data enters frame.
fiund's sourcing angle
Agentic training data barely exists as a licensable category yet. fiund treats it as a sourcing brief from day one. We source to a brief and clear the rights before anything moves, so what you receive is both useful and defensible in diligence.
Rights posture
Signed licence, explicit training rights, separate voice/likeness consent, nothing scraped. See the rights & provenance guides.
Frequently asked questions
How is this different from robotics trajectory data?
Same episode structure, different world: digital trajectories capture screen state and input events instead of pose and contact, and collection is software instrumentation rather than cameras. World-model use stresses long horizons and causal alignment harder than policy training does.
Why does environment versioning matter so much?
Because a UI is the world being modelled — a silent app update changes the dynamics mid-corpus, and the model learns an inconsistent environment. Pinned versions per episode let you train coherently and measure generalization across versions deliberately.
Other data for World models
More Agentic trajectories use cases
Need Agentic trajectories for World models?
Whether you're building a model or sitting on an archive, the first conversation is short and specific.
Send a brief