Machine intelligence
Can agentic systems become dependable enough to own long-running technical work?
We are interested in evaluation, memory, oversight, human–model interfaces and the systems around the model—not another thin wrapper.
New lab / forming now
Duskfields is a new independent lab working across machine intelligence, scientific tooling and computing infrastructure. We are setting it up now.
Read the honest version↓State of the lab
Updated / Sep 2026The honest version
There is no portfolio to browse yet.
Right now, we are choosing the first research programmes, building shared technical infrastructure, and speaking with people who know where the real problems are.
We will update this site when there is work worth showing—not to manufacture the appearance of momentum.
Initial search areas
Not product announcementsThese are starting coordinates, not a mandate. A good project can change the map.
Can agentic systems become dependable enough to own long-running technical work?
We are interested in evaluation, memory, oversight, human–model interfaces and the systems around the model—not another thin wrapper.
What would materially increase the rate at which a small research team can learn?
Software for instrument control, simulation, experimental design, evidence tracking and reproducible technical work.
Which parts of the emerging stack are still missing, fragile or needlessly expensive?
Infrastructure that makes new forms of computation usable in real organisations under real constraints.
Method
A simple filterWe expect most ideas to stop. The useful ones become clearer under pressure.
Write down what must be true, what would prove us wrong, and why the answer matters.
A prototype should carry actual load. Demos designed only to impress do not count.
Performance, cost, reliability and behaviour—not impressions, engagement theatre or vibes.
Publish the result, continue the programme, or form a company when the evidence supports it.
Working rule / 001
Contact
Channel opening soonIf you are working on an unusual technical system, have research that deserves a path into the world, or know where an important tool is missing, we would like to hear it.
Short, concrete emails are best.