What does agent training data look like?
It records work as a sequence of actions, not as a single question and answer. τ-bench, a research benchmark published in 2024, gives an agent "domain-specific API tools and policy guidelines", puts it in a conversation with a simulated user and scores it by comparing "the database state at the end of a conversation with the annotated goal state" (Yao et al., τ-bench, arXiv). SWE-bench does the same for code, using real GitHub issues and the pull requests that resolved them (SWE-bench, arXiv).
Why do SOPs, documents and code make good agent training data?
They are the policies, tasks and outcomes agents are tested against. A written SOP plays the part of τ-bench's policy guidelines; a support ticket or change request is a task with a known end state; a merged commit is a recorded solution. An AI training startup's chief revenue officer told Business Insider that AI developers increasingly want "long-horizon" data that can provide a full picture of how work gets done (Business Insider, Oct 2026), and MIT Technology Review described one buyer building "reinforcement-learning gyms" where its clients' agents practise tasks (MIT Technology Review, Oct 13 2026).
What does agent training data mean for what you get paid and control?
It favours records that link steps across systems, and most companies already keep those links in their ticketing, documentation and code tools. Much of a company's data already lives in third-party software. Licensing it means re-licensing what already exists, and getting paid for it rather than giving it away for free. Linked records also carry more personal detail than any single document, because they follow people from one tool to the next. fearofgod.ai de-identifies the data before it goes to buyers. Raw data is deleted according to a schedule agreed with the partner. Your company can be paid more than once, through renewals and repeat licences.