AI agents with safeguards
Multi-agent systems that plan, delegate and check their own work, with hard limits on money, credentials and outbound actions.
Read more →Software · Data · AI
Software, data and AI that work correctly without supervision.
We build and test systems that have to be right when nobody is watching: AI agents, machine-learning models, and coordination between independent parties. Our rule is simple: don't trust, verify.
What we do
Model validation
Independent testing of machine-learning models: does the result hold on unseen data, under real costs and delays, against a random baseline?
Read more →Coordination research
How agents and systems owned by different parties agree on who does what, without a referee who works for one side.
Read more →How we work
Assume it is wrong until tested.
A good result is a reason to look harder. Every model, agent and claim goes through the same checks before anything builds on it.
Our approach →- 1Lock the final test. Some data stays untouched until the very end, and is used once.
- 2Use real-world conditions. Real costs, real delays, real limits.
- 3Keep people on the decisions that matter. Money, access and anything public need a human yes.
Have a question about one of these?
Tell us what you are working on. A few lines is enough.