Scattered Intelligence is an applied research lab building training environments from real work.
Models can now solve research problems that have stood open for decades, yet still stumble on the everyday work of a company. What they lack is data: the reasoning, decisions and changing state of an organisation, captured as the work happens.
Software keeps the result. The reasoning is lost.
A CRM records that a deal changed stage. It does not record what people read, weighed and decided along the way. The judgment that runs a company is scattered across people, tools and time. We capture it.
- 1.1
Capture
Agents working inside real companies leave a trace every time they act, with a person approving, correcting or rejecting the work.
- 1.2
Reconstruct
We rebuild the environment around each trace: the documents, the state of the tools, the intent behind the request and what happened next.
- 1.3
Verify
Every task ships with a verifier grounded in what actually happened, and every run is checked for reward hacking before it leaves us.
All knowledge work is research.
- A-01Tasks
Long-horizon knowledge work
Hard tasks drawn from inside companies: many documents, many tools, goals that take an agent days or weeks to reach rather than minutes.
- A-02Environments
Environments from live work
Offline snapshots of the environment around real agent traces, including what the user intended and what they did after the output.
- A-03Verifiers
Grading open-ended work
Knowledge work has many good answers and many bad paths to them. Rubrics and evals informed by real outcomes, not guesses.
Build the data frontier models learn from.
We are early. We want to meet researchers who have built environments and verifiers, and engineers who can put agents to work inside real companies.
Correspondence
Research labs, operators and collaborators.