Real company data for RL environments

Short version

An RL environment is a sealed copy of a working world, such as a company's files, inbox, database and spreadsheets, plus tasks and automatic checks. A researcher drops an agent in, gets a score, and runs thousands of copies to train it. Environments built on synthetic companies are easy to make and easy to beat. Environments built on the records of real companies are harder to fake and closer to the work your users want done. Grepped supplies that data.

What goes into an RL environment

Most environments for agent training have four parts:

1. The world. A file system, mailbox, CRM, ERP or code base the agent can explore.
2. Tools. Search files, open a spreadsheet, query a database, send an email, run a command.
3. Tasks. Prompts like "reconcile March's invoices against the purchase order log and flag the mismatches".
4. Verifiers. Code or rubrics that check the result automatically.

Packaged together, usually in a container, the same environment works as an eval and as a training ground.

Why real company data makes better environments

The mess is the hard part

Real companies have half-filled spreadsheets, three versions of the same quote, folders named "final_v2_USE THIS" and email threads that change direction halfway. That's what makes back-office work hard for agents, and it's very hard to generate convincingly.

The answers are already recorded

A company archive shows what actually happened: which quote was sent, which invoice was paid, which ticket was closed and how. Those outcomes become ground truth for verifiers, so you're not relying on a model or an annotator to guess the right answer.

Long horizons come free

A single job at a construction firm can run from enquiry to estimate, change orders, invoicing and a dispute over months, across five tools. That's the long-horizon, multi-tool work current models fail at, and it's already sitting in the records.

Your product logs can't cover it

Traces from your own users only show what people already try with today's models. Real company archives cover work nobody is asking agents to do yet.

What's available

The Grepped marketplace lists archives from real companies across:

• Construction and engineering: estimates, change orders, job costing, drawings, CAD and BIM.
• Manufacturing and industrial: job folders, quality and root-cause records, formulations, maintenance logs.
• Finance and accounting: workbooks, close processes, tax cases, lending and underwriting.
• Software companies: repositories with pull requests and reviews, issues, support tickets, Slack and product docs, including wound-down startups.
• Services and retail: CRM, scheduling, support conversations, POS and ERP data from operating businesses.
• Healthcare operations, legal and science, where buyer rules and permissions are checked per listing.

Every listing shows the domain, the tools the records came from, the size of the team behind them and whether the data has been sold before.

What to check before you buy

Most failed data deals fail on quality, not price. Before you commit, ask:

• Is it real? Records from an operating company, not generated or reconstructed.
• Is it contaminated? Check whether any of it is on the open web and likely in pre-training corpora.
• Can tasks be verified? Look for recorded outcomes you can check an agent against.
• Are the rights clear? Who owns the records, whether client contracts restrict them, and whether anyone holds exclusivity.
• Does the sample match the listing? Open it, run an agent on a few tasks, and look for broken files or gaps.

How buying works

1. Join Buyside. Buyside is where qualified buyers browse listings. We let new buyers in from the waitlist.
2. Pull a sample. Samples are covered by our mutual NDA and download in about a minute.
3. Ask for a quote. Our team comes back with a price and the licence terms. Whether a deal is exclusive depends on the owner and the listing.

Can't find what you need? Tell us what you're looking for and we'll source it.

Common questions

Do you deliver ready-made environments or raw data?

Raw archives are the starting point. We can package them closer to what your team trains on, and we work with environment builders who turn archives into tasks and verifiers.

Can I see data before signing anything big?

Yes. You sign our mutual NDA, then pull samples before any licence.

Can I request a specific industry or tool?

Yes. Tell us the domain, tools, history and team size you need and we'll look for it.