What makes a dataset agent-ready

"Agent-ready" is turning into a common label for training data, and it is used loosely. A dataset is not agent-ready because it is large, recent or full of workplace vocabulary. It is agent-ready when an agent can learn from it how work actually gets done, and when a lab can use it without legal or privacy surprises. Those are two separate tests, and a good dataset passes both.

Text-heavy datasets were judged mostly on volume and quality of language. Agents are judged on whether they complete tasks: reading state, choosing actions, using tools, coordinating with people and reaching a correct outcome. Data that teaches those skills has to preserve the structure of the work, not just the words.

That changes what matters. A million disconnected messages may teach very little about how a project moves from request to delivery. A few years of one company's connected records, by contrast, can contain thousands of complete examples. We explain how that affects value in how company data is valued.

Looking across the data labs ask for, the same properties come up repeatedly. An agent-ready dataset usually has:

• Several connected systems. Chat, email, docs, tickets, code, CRM, support and finance from the same company, linked to each other.
• Years of history. Enough time for long projects to finish, for decisions to show their consequences, and for patterns to repeat.
• Complete threads. Conversations, tickets and reviews that run from start to finish, including the dead ends.
• Consistent identity. Every person, company and account carries the same token in every tool, so ownership and handoffs are visible.
• Preserved structure. Statuses, assignees, timestamps, comments, diffs and links kept in their native form rather than flattened into plain text.
• Visible outcomes. Tickets closed or reopened, deals won or lost, releases shipped or rolled back.
• A realistic mix of work. Not only engineering, but also the sales, support, finance and operations tasks agents are increasingly asked to do.

Few datasets are perfect on every point. But a source missing most of them will struggle to support either training environments or evaluations, however large it is.

Much of the value of operational data sits in metadata that is easy to discard. A ticket's change history shows who moved it, when and why. A pull request's review comments show what the reviewer cared about. Timestamps reveal how long things took and in what order they happened.

When data is exported carelessly, or anonymized by flattening everything into text and scrubbing it, that structure disappears. The result reads fine to a person skimming it and is much weaker for training. Keeping native structure, and anonymizing in a way that respects it, is a large part of what separates an agent-ready dataset from a pile of logs. For a concrete picture, see what connected data looks like in practice.

The second test is whether the data can be used responsibly. Here the techniques matter as much as the intent.

Consistent tokenization replaces every person, company, address, email, phone number and account with a pseudonym that stays the same across all tools and all years. PERSON_014 is the same person in Slack, email and Jira. That keeps the work coherent while removing identity. Credentials, API keys, card numbers and government IDs are removed outright. Faces, logos and signatures are blacked out in images and PDFs. HR, payroll, medical and legally privileged folders are excluded by default, and direct messages stay out unless the company includes them and has the right to.

A dataset anonymized this way is safer and more useful at once. A dataset where names were simply deleted is often neither: identifying context remains, and the links that made the data valuable are gone.

An agent-ready dataset also has a clean answer to the question of where it came from and what it may be used for. Labs increasingly ask for this as a matter of course. See data provenance is becoming a buying requirement.

In practice that means:

• The data was licensed directly from the company, signed by someone with authority to do so
• Scope and exclusions were agreed and annexed to the contract before any data moved
• The licence restricts use to training and evaluation, with no marketing, profiling, contact or re-identification
• No third party touched the raw data, and the raw archive was deleted within 30 days of sign-off, with a deletion certificate

Without this, a dataset that is technically excellent can still be unusable.

Most agent-ready data still sits inside companies that have never thought of it as an asset. Some are operating businesses; others are winding down and about to cancel the systems that hold it. Bringing it to market takes outbound work: finding companies with the right tool stack and history, and reaching the person who can sign.

That is the approach we take. Labs send a sourcing brief with sector, headcount, tool stack, years of history, regions, and operating or winding-down status. We build a target list from company and technographic data, our onshore reps call the decision-makers, and each archive is evaluated, anonymized in-house and delivered on a rolling basis.

If you are looking for data that passes both tests, send us a sourcing brief.