Knowledge Work
Make one decision from sources that disagree
A product manager pieces a decision together from threads and docs. A financial analyst reads a stack of reports and writes a note. We build environments around real job functions.
RL environments for knowledge work agents
RL environments for knowledge work agents recreate a job. They hold the documents, tools and messages a professional works with, and a task that ends in a deliverable. The agent has to decide what matters, find it and produce work a practitioner would accept.
We build each environment with practitioners from the job, who ground the case, the documents and the checks in how the work is really done. Experts who didn’t build a task review it before it ships.
Finance is part of this domain. FinanceBench, our open-book benchmark of 10,231 questions about public companies, is used by post-training teams as an environment. Where a task has a right answer in the source records, the reward checks the answer against those records.
Real questions start vague. In DETOUR, an agent had to identify something a person could only partly describe, over several turns, and the best model tested got 36% across text, image, audio and video.
Published research
DETOUR
A two-agent benchmark for tip-of-the-tongue search across several turns.
FinanceBench
Open-book financial question answering about publicly traded companies.
Generative Simulators
Environments that generate their own tasks, world dynamics and rewards.
Questions about knowledge work agents
How is a knowledge work environment different from a computer use environment?
Computer use is about operating the software. Knowledge work is about the result, such as the memo, the decision or the reconciled account. An agent needs the first to do the second, and we build both.
Which job functions do you build for?
We build around real job functions, from product managers to financial analysts. Each environment carries the tools and documents of that job, so an agent has to adapt to whatever the work puts in front of it.
Who designs the tasks?
Practitioners from the job, working with our researchers. Experts write the case and the checks, and separate experts review them. Models help with scale, and experts make the final call.
Can rewards be checked by code instead of a model judge?
Often, yes. When the answer lives in source records, such as a filing or a ledger, we check it against them. Some work has no single right answer. There we use criteria written by practitioners.
Is the data in your environments real?
Each environment starts from real-world material where we can license it, and experts build on top of it. Private details are made up for the task, and example documents are fictional.