Buy agent trajectory data and RL environments
Last updated How we research
In short
Agent trajectory data records each step an AI agent or a human expert takes to complete a task: the observations, clicks, tool calls and the result. Labs use it to train and evaluate agents, often alongside a replayable environment. DataVendor reports that post-training work now creates stronger demand for evals, environments and real workflow traces.
Key takeaways
- AI teams use agent trajectory data for (1) supervised fine-tuning of computer-use and tool-using agents, (2) reinforcement learning with replayable environments and (3) reward model and verifier training on successful and failed runs.
- Usually priced per task or trajectory, or per environment with its taskset.
- Common delivery formats: JSONL step logs with observation, action and result, screenshots or screen recordings paired with action logs and containerized environments such as Docker images.
- Check before you license agent trajectory data: (1) each trajectory has a verified outcome, (2) environment replays reproduce the recorded result and (3) US: personal information in screenshots and observations removed under the CCPA and other state privacy laws.
- Every seller we shortlist passes 8 vetting checks before you see a sample.
Agent trajectory data: buyer's spec
- Used for
- Supervised fine-tuning of computer-use and tool-using agents
- Reinforcement learning with replayable environments
- Reward model and verifier training on successful and failed runs
- Evaluation of agents on held-out tasks
- Delivery formats
- JSONL step logs with observation, action and result
- Screenshots or screen recordings paired with action logs
- Containerized environments such as Docker images
- Task definitions with scoring functions
- Usually priced
- Per task or trajectory, or per environment with its taskset
- Check before you sign
- Each trajectory has a verified outcome
- Environment replays reproduce the recorded result
- US: personal information in screenshots and observations removed under the CCPA and other state privacy laws
- Rights to the software inside the environment
- Split between human and model-generated runs disclosed
- No overlap with public benchmarks
Why AI teams work with us
- 1
Every seller of agent trajectory data we shortlist has passed eight vetting checks, from authority to sign through to a re-identification test.
- 2
Each shortlist comes with a vetting file per seller: rights evidence, provenance record, de-identification method, data manifest and proposed license terms.
- 3
Samples are drawn at random across the archive and labeled with the sampling method, so you can judge the whole package rather than its best rows.
- 4
You contract directly with the seller, and we help agree acceptance tests, staged payments and remedies before any data is delivered.
What is agent trajectory data?
Agent trajectory data is a step-by-step log of an agent completing a workflow, linking each action to what the agent saw and what happened next. DataVendor defines trajectories as records of the steps an AI agent takes while completing a workflow, and notes that they connect model actions to measurable outcomes.
A trajectory can come from a model run or from a person doing the task while every action is logged. OpenAI's 2022 InstructGPT paper is an early, well-documented example of training on human demonstrations, although its demonstrations were written answers rather than tool use. The glossary defines an agent trajectory in brief.
What is the difference between a trajectory and an RL environment?
A trajectory is a recording, and an RL environment is the sandbox in which the task can be run again. DataVendor describes replayable environments as the software or sandbox needed to run the task again, and on its marketplace environments cannot be listed on their own, only as tasksets with their environments.
| Asset | What it contains | Main use |
|---|---|---|
| Trajectory | Observations, actions and outcome for one run | Supervised fine-tuning and reward modeling |
| Taskset | Task prompts with scoring rules or expected results | Evaluation and RL rewards |
| RL environment | Runnable software or sandbox for the tasks | Reinforcement learning at scale |
Our guide to agent trajectories and RL environments covers each asset in more depth, and the glossary defines an RL environment.
Why is demand for agent trajectory data rising?
Demand for agent trajectory data is rising because post-training has become a main route to better agents. DataVendor states that post-training work now creates stronger demand for evals, environments and real workflow traces.
Real company archives feed this too. MIT Technology Review reported in October 2026 that Surge AI buys company data to build reinforcement learning gyms where clients' agents practice tasks such as debugging complex IT architecture or drafting appellate briefs. The same article noted that no company has published tests comparing real-world and alternative training data, so the size of the gain is still unclear.
Where does agent trajectory data come from?
Agent trajectory data can come from logs of AI products where agents already act for users, from expert demonstrations recorded for the purpose, or from company records rebuilt into tasks. MIT Technology Review reported that buyers value a codebase by estimating how many useful tasks it might yield, and the same reasoning applies to ticket queues and approval workflows.
Product logs need care. Prompt and response logs, agent run traces and user ratings all describe agent behavior, but they belong to the product's users as much as to the company, so they need a privacy notice that covers the use or proper de-identification before licensing. California residents can also opt out of the sale or sharing of their personal information under the CCPA, according to the California Attorney General.
How is agent trajectory data priced?
Agent trajectory data is priced per task, per trajectory or as a bundle of environment and taskset, depending on the seller. There is no public price index, and DataVendor warns that a listed price or a large catalog does not establish what an asset will earn.
Ask whether a quote covers failed runs as well as successful ones. Failed runs with a recorded reason can be useful for training verifiers, so check whether the seller kept them. Our guide to AI training data pricing models compares per-unit and flat-fee structures.
What should you check before buying agent trajectories?
Before buying agent trajectories, replay a random sample in the environment and confirm the recorded outcome. A trajectory that does not reproduce cannot be trusted as a training signal.
- Outcome verification method, automatic or human
- Share of trajectories produced by models rather than people
- Personal data in screenshots, form fields and observations
- Licenses for any commercial software inside the environment
- Overlap with public benchmarks that would contaminate evaluation
To source vetted trajectory data and environments, book a free consultation. Our due diligence checklist covers the contract side, and evaluation datasets explains how to keep test tasks clean.
Frequently asked questions
What does an agent trajectory contain?
Can company workflows be turned into agent training data?
Are human or model-generated trajectories better?
Can an RL environment be bought on its own?
Sources
- AI training data marketplaces, DataVendor, accessed November 2026
- AI labs are buying data from dead startups to train their models, MIT Technology Review, accessed November 2026
- Training language models to follow instructions with human feedback, arXiv (Ouyang et al., 2022), accessed November 2026
- California Consumer Privacy Act (CCPA), State of California Department of Justice, Office of the Attorney General, accessed November 2026
For AI teams
Ready to source?
Book a consultation to brief us. We shortlist vetted sellers, each with a vetting file covering rights evidence, provenance and a representative sample.
Related guides
- Agent trajectories and RL environments explained
Data types
- How to buy AI training data
Buying data
- AI training data pricing models
Pricing and value
- AI training data due diligence checklist
Buying data
- How to evaluate a dataset sample before you buy
Buying data