# Datalab Document Agent Beta > Datalab Document Agent is now in closed beta: specify what you want, provide sample documents, iterate on what correct means, and our agent will devise a custom suite of verifiers and tools built on our state of the art models and pipelines to produce the right document workflow for your needs. - Canonical: https://www.datalab.to/blog/document-agent-beta - Published: 2026-08-24 - Authors: Ashish Uppala, Tarun Menta Over the last 2 years, we’ve parsed an incredible diversity of documents with our leading models Surya, Marker, and now Chandra. And in conversation with customers over that time, we’ve come to realize a few things: - Document diversity is endless (you try parsing a scanned document with ink bleeding through from the reverse side) - Parsing is just the beginning of a typically complicated document transformation / workflow - Customer preferences add an extra layer of combinatorial complexity on parsing properly What does it mean to parse a document *correctly*? Sometimes it’s simple — did you literally represent information accurately relative to how it was on the page? Other times, a user will have opinions: “well… yes… but could you do it like this instead?” Indeed, there are objective and subjective notions of correctness when you root conversion in a customer’s perspective. As we’ve circled around this problem over the last year, we’ve realized that we need a general solution to what feels like an infinitely complex problem. Document Agent, now in closed beta, is our answer. ## Range of a Conversion A parse is just the start. When you convert a PDF, DOCX, Spreadsheet, etc. you need to do something to it or get something out of it. These use cases can range from: 1. Providing a schema with key fields to pull information out, the classic structured extraction use case 2. Remediating a PDF so it meets accessibility standards (WCAG compliance) 3. Converting a PDF into a JATS-XML format, if you’re a scholarly publisher 4. Doing a conversion but getting redlines out of the document 5. Doing a conversion, but formatting lists in a really specific way 6. Doing a conversion, but oh it’s a gnarly clinical trial and the tables are super dense and humans can visually identify where the columns are but no one really drew lines and it’s kind of ambiguous And plenty more. In the past this might have required highly custom workflows and engineering effort on a customer’s side, but when we have: - agentic capabilities and frontier intelligence as a resource - state of the art conversion and extraction models and pipelines internally at Datalab We wondered if we could build an agent harness that leveraged our core conversion primitives to streamline how we solve these problems for customers. ## Dogfooding our agent We recently wrote about our [new processor for making PDFs WCAG compliant](/blog/wcag-accessible-pdfs), and our other one to convert [PDFs into JATS-XML for scholarly publishers](/blog/pdf-to-jats-xml). Internally, we used our own document agent and started a new design session, iterating with it by specifying our intent, providing samples of documents, and creating verification criteria around what a correct conversion looked like. A Document Agent design session. On the left, a chat where the user describes converting scholarly PDFs into JATS XML with resolved DOI citations and multi-article handling; on the right, the proposed contract — a goal, four rules, and pass/fail checks marked blocking or advisory — committed as v0.1. *A design session from building our JATS processor: intent on the left; the versioned contract — goal, rules, and checks — on the right.* Our agent leveraged internal models and other pipelines trained by our research team as tools it could use and spun on the problem by devising different strategies, checking results against a concrete set of verifiers it designed with us, all until it figured out what worked reliably. This looks a lot like how a motivated person with a problem might approach it: they know what they want done and are armed with examples and definitions of success, but they have to coordinate across a team internally that stitches various services, maintains infra, etc. Our agent collapses that loop dramatically, giving anyone the power to get what they need from their documents. As it worked and as quality on our benchmarks improved, these internal agent explorations were saved and templatized into our now standard JATS and WCAG processors, ready to be used directly with a simple API call! ## Closed Beta We’re really excited by the potential this approach has — it can feel daunting and cumbersome to work on some of these problems and instead, in our experience, it’s becoming magical. We’ll be releasing templates regularly, but access to the full custom agent is still in closed beta. If you’re dealing with complicated document workflows, reach out to us at [support@datalab.to](mailto:support@datalab.to) to get access — we’re stoked to see what you build with it!