Agent trajectory data and RL environments
Recorded task sessions (screens, actions, outcomes) and replicas of real software environments that agents can practice in.
- Measured in
- Trajectories or tasks completed, applications covered, average steps per task.
- Who buys
- Foundation model labs, agent builders and vendors that build RL environments for labs.
- Where it comes from
- Companies with skilled operators in specific software, workspace exports from closed companies, and purpose-built collection.
- Rights to clear
- Software vendor terms, employee consent, client data visible on screens.
- Privacy
- Screens show customer and colleague data; masking and synthetic substitution are standard.
- Pricing signal
- New trajectories are usually priced per task. One startup had about $15,000 pending for its workspace data (MIT Technology Review, 2026).
- Agent trajectories are step-by-step recordings of a person completing a task in software: every click, keystroke, screen and result.
- Labs training computer-use and workflow agents buy trajectories and the environments to practice in, because the web does not contain them.
- MIT Technology Review reported Surge AI buying data for reinforcement-learning environments where agents practice tasks.
- Workspace exports from tools like Google Workspace, Notion and Figma are being bought to build realistic practice environments.
Why agents need trajectories
An agent that books freight, reconciles invoices or files insurance claims has to operate the same software people use. Text alone cannot teach that. Recordings of skilled people doing real tasks, with the outcome attached, show the agent what good looks like, and practice environments let it learn by trial.
The market
MIT Technology Review reported that buyers of shut-down startup data include vendors building simulated workspaces for AI training, and that Surge AI buys data for reinforcement-learning environments (MIT Technology Review). Exports from Google Workspace, Notion, Figma and calendars were among the data types described.
Two kinds of supply
- Historic workspace exports. Documents, tickets, calendars and messages from a real company, used to build a realistic environment. These need the same rights and privacy work as email and workplace chat.
- New recorded trajectories. People with domain skill complete defined tasks while their sessions are recorded. This is collection work, so the deal is a services agreement as well as a license.
Quality markers
Clear task definitions, success criteria, realistic data in the environment, diverse operators and recorded failures and recoveries.
Questions
Can our operations team earn money by recording tasks?
Possibly. Some buyers fund recorded task collection in specific software. Staff consent and client data on screen are the main issues to plan for.
What is an RL environment?
A safe copy of a software setting where an AI agent can attempt tasks repeatedly and be scored, so it learns from trial and error. See the glossary entry.
Related guides
Related data types
Sources
Need agent trajectory data, or hold it?
Tell us what you need or what you hold. We check rights and privacy first, then introduce vetted counterparties. You license directly with each other.
- We reply to every enquiry within two working days.
- No upfront fees for sellers
- We never hold your raw data