Computer-Use and Agent Trajectory Data: What It Is and Who Buys It

Computer-use and agent trajectory data is the step-by-step record of an AI agent or a person completing a task on a computer: the prompt, each action or tool call, what the screen or system returned, and the outcome. micro1 states USD 1 to 2 million a year and Mercor USD 20,000 to 480,000 per deal for it (checked 2026-11-09), and Magma handles it for holders.

The page treats agent trajectories as one data type in AI training data licensing: where the records come from, who asks for them and whose rights sit inside them. DataPacts takes no cut of any license.

Every figure and term below is a claim by the party named, with its check date, not a paid amount; the approach is explained under how we source, date and check every figure.

What Is Computer-Use and Agent Trajectory Data?

A trajectory is a complete, ordered log of one task: what the agent was asked, what it did at each step, what came back, and how the task ended. Mercor names the parts it wants in a session log as prompts, responses and every tool call (checked 2026-11-09).

The table compares seven components of a trajectory with the acquirer that names each (as of 2026-11-10; program pages checked 2026-11-09).

The term "computer-use agent" also names a software product that operates a screen, a sense this page does not cover. A trajectory also differs from a static dataset because it records actions and their results, not only content.

Where Trajectories Come From: Production Agents, Recorded Work, Developer Tools and Simulated Environments

Trajectories come from four places: AI products and agents running for real users, people recorded doing multi-step work, developer tools that log coding sessions, and simulated environments built for training.

The table compares the four sources by holder, example on record and main rights question (as of 2026-11-10).

Gray Swan states more than 1,000,000 attack trajectories from its Arena (2026-05-28), and Datacurve lists agent trajectories among its products (checked 2026-11-09); both produce trajectories rather than license them from holders. Because trajectories recorded in RL environments are sold with the environment, that topic has its own page. GitHub states that it trains on Copilot Free, Pro and Pro+ interaction data from 2026-04-24.

Who Licenses Agent Trajectories: Programs, a Marketplace and Lab Agreements

Eight organizations on record acquire, handle or produce trajectory data: two company data programs, one marketplace, two vendor programs, one lab agreement and two builders that produce their own.

The table lists the eight alphabetically with their role, what each asks for or holds, and the check date (as of 2026-11-10; programs checked 2026-11-09).

micro1: Recurring Licenses for Approved Agent Trajectories

micro1 states that it licenses qualifying agent trajectories for USD 1 to 2 million a year with the company's approval, after an optional agent build-out in months 1 and 2, checked 2026-11-09. Its main page says "We pay in full within 30 days of acceptance", while its FAQ describes up to 30 days of review, then an invoice, with payment within 30 days of invoicing.

Eligibility starts at 30 full-time staff, with data primarily in English and the US prioritized, and micro1 licenses the data only to buyers in the US and allied countries. The full terms of micro1 as a data buyer sit on its profile, and micro1 states that it de-identifies the data itself.

The Data License Master Terms (last updated 2026-11-08) name agent trajectories among the licensed data. Their grant is perpetual, irrevocable and sublicensable.

Mercor: Session Logs From AI Products in Production

Mercor's listing for AI product companies, published 2026-10-17, asks for complete session logs from a product in production for at least six months, and Mercor states a deal range of USD 20,000 to 480,000 to the company, checked 2026-11-09. The listing is open to companies of 5 to 200 employees.

The listing requires prompts, responses and every tool call, plus user terms that allow licensing of de-identified usage data or a way to get consent, and it names tools such as Claude Code, Codex, OpenHands, Cline, OpenRouter, Langfuse and LangSmith. The listing is one of 25 covered under Mercor as a data buyer.

Mercor states three cases that are not a fit:

• Synthetic or scripted traffic
• Logs that cannot be stripped of personal information
• Truncated or partial traces

Magma: A Marketplace That Licenses Agent Traces for a Revenue Share

Magma Intelligence Inc. runs a marketplace on which companies building AI agents license agent traces, corrections and completed workflows to AI labs for a revenue share whose percentage is set in onboarding and not published. Magma handles the data on the holder's behalf.

The process starts with a two-minute form and a 20-minute onboarding call, then a read-only encrypted connection limited to the approved scope, after which Magma packages the data for labs. Other AI training data marketplaces are compared on fees and eligibility.

Magma's terms (last updated 2026-07-28) require a separate written definitive agreement, allow regulated data such as HIPAA data only under a separate agreement, cap the site's liability at USD 100 and apply Texas law.

Labs and Vendors: The Cursor Agreement, Labelbox Prism and SuperAnnotate

Two vendor programs name trajectory-like records among the data they license, and one lab agreement covers developer interaction data from a coding tool. Labelbox Prism wants code generation and edit traces and cross-system interaction data, and states that licenses are exclusive by default, with a kill clause forcing deletion within 30 days and models already trained that cannot be unlearned (checked 2026-11-09).

SuperAnnotate lists screen recordings of multi-step tasks and states that "You're licensing rights for a set period, not selling the data", with the seller approving each buyer (checked 2026-11-09).

The 2026-04-19 agreement between SpaceX (including xAI) and Anysphere, Inc. (Cursor) is a compute and option agreement that covers coding prompts and iteration cycles, and the SEC filing sets no data price; see the Cursor and SpaceX AI developer interaction data agreement. The Grok 4.7 training-content summary (v1, 2026-09-21) lists anonymized Cursor workflow data, according to SpaceXAI.

What Buyers Check Before Accepting Trajectory Data

Buyers check four things before they accept trajectories: that the logs are complete, that they come from real use, that personal data can be removed, and that users or the company allowed the license.

The table compares what Mercor, micro1 and Magma state on eight requirements (as of 2026-11-10; checked 2026-11-09).

Only Mercor states tests on real use and completeness. The size, country and data-type filters are compared under buyers' published acceptance criteria.

Whose Data Is Inside a Trajectory: User Terms, Privacy Notices and De-identification

A trajectory from a live product carries other people's data: the user's prompts, the files and screens the agent touched, and any customer or third-party records those screens showed. Mercor therefore asks for user terms that allow licensing of de-identified usage data, or a way to get consent (checked 2026-11-09).

Regulators attach consequences to terms that change after the data is collected. FTC staff wrote on 2024-02-13 that it may be unfair or deceptive for a company to adopt more permissive data practices, such as AI training through a surreptitious, retroactive amendment to its terms or privacy policy, and on 2024-01-09 that prior FTC actions required deletion of models and algorithms built on unlawfully obtained data; both are staff posts, not rules. Connecticut P.A. 25-113 requires privacy notices to state whether personal data is collected, used or sold to train large language models from 2026-07-01, and the COPPA Rule amendments (90 FR 16918, effective 2025-06-23, compliance 2026-04-22) require separate parental consent for disclosures to train AI (90 FR 16950).

Removal of personal data is the seller's task unless the acquirer states otherwise, as micro1 does. The steps are covered under de-identification before licensing data. Regulation (EU) 2016/679 has applied since 2018-05-25, and the EDPB's Guidelines 02/2026 on anonymisation (version 1.0, adopted 2026-07-07) are a consultation draft, not final as of 2026-11-09. This page is information, not legal advice.

Stated Prices for Trajectory Licenses, by Stating Party and Date

No market price for trajectory data exists; the figures on record are two acquirers' claims for trajectories, two vendors' program-wide ranges and one revenue share without a percentage. Each is a claim with a check date, not a paid amount.

The table lists every stated figure with its stating party, scope and date (as of 2026-11-10).

No acquirer states a price per trajectory, per session or per step, so the figures are not averaged or converted. Rates for other data types, mostly per unit, appear under AI training data prices.

Licensing Agent Trajectories: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

Licensing trajectories can turn logs an AI product already keeps into recurring income, but the license usually outlives any later change of mind.

Benefits

• Recurring income: micro1 states recurring licenses from launch (checked 2026-11-09).
• Existing logs: Mercor asks for session logs of a product that already runs, including logs kept in tools such as Langfuse and LangSmith.
• Control over scope: micro1 licenses with the company's approval, and SuperAnnotate lets the seller approve each buyer and block competitors permanently.

Risks

• User and customer data sit inside the logs, and a retroactive terms change may be unfair or deceptive according to FTC staff (2024-02-13).
• A regulator can require deletion of models built on unlawfully obtained data, as FTC staff described on 2024-01-09.
• Screens an agent touched can show customers' confidential workflows and third-party records.

Downsides

• micro1's master terms grant a perpetual, irrevocable license, and raw data is deleted within 120 days but with no duty to delete or retrain models built on it (sections 1.2 and 4.4).
• Prism states that models already trained cannot be unlearned, even under its 30-day kill clause.
• De-identification and approval fall on the seller, and Mercor rejects truncated or unstrippable logs.

Alternatives

• Keep trajectories for internal evaluation.
• License only through a scoped, read-only connection, as Magma describes.
• Sell an environment or task set instead of the logs.
• Not licensing at all: whether to sell or license your data is decided there.

None of the options is preferred; each depends on the product, its users' terms and the acquirer's license. A company that declines every offer keeps its logs and its freedom to change course.

Where Trajectory Data Meets RL Environments, Code, Programs and the Seller Guide

Trajectory data sits between three neighbors: the environments that generate trajectories, the codebases behind coding agents, and the company data programs that license operating records.

• Every data type AI companies buy: the parent index.
• Source code and codebases behind coding agents: code is not developed here.
• Company data partnership programs that license operating records.
• Every buyer and program, compared without fees.
• How to sell an RL environment or task set instead of logs.
• The complete seller guide.
• Program terms compared, dated.
• RL environments for computer-use agents: builders of simulated software.
• Magma's marketplace terms, Labelbox Prism program terms and SuperAnnotate program terms.
• Software companies licensing code and product data.

Frequently Asked Questions

Is an agent trajectory the same as RLHF data?

No, a trajectory logs the actions of one task run, while RLHF data is people ranking or rating model outputs. The two are collected and licensed differently. The ranking side is covered under what RLHF data needs and who supplies it.

Can individuals get paid to record their screen for AI training?

This page covers companies licensing their records; individual paid work is a separate market. The work is paid per task or hour, not licensed as a company asset. Contributor pay is compared under AI training jobs and platforms compared by pay.

Do coding tools train on their own users' sessions?

Some state that they do. GitHub states it trains on Copilot Free, Pro and Pro+ interaction data from 2026-04-24, and the Grok 4.7 summary (2026-09-21) names anonymized Cursor workflow data. Policies are compared under when services train on their users.

Can a company license trajectories its users never agreed to share?

Mercor requires user terms that allow licensing of de-identified usage data, or a way to get consent (checked 2026-11-09). FTC staff warned on 2024-02-13 against retroactive terms changes. The consent step is covered under customer consent and confidentiality clauses.

Do buyers accept synthetic trajectories?

Mercor does not, as it lists synthetic or scripted traffic as not a fit. The listing was checked 2026-11-09. Synthetic data is a separate market with its own acquirers, described under synthetic data for AI training and its limits.