Democratising the development of safe agentic AI

We're an agentic AI applied research lab, building on the frontier of multi-agent software systems.

Our research is underpinned by the following technical principles:

  • Building performant AI agents is a design problem, not a technical problem. The bottleneck isn't compute or infrastructure — it's discovering the right design.
  • Agents are designs to be discovered, not programs to be written.
  • Reliability comes from composition, not inheritance — explicit structure over hidden state.
  • Transparency and observability are key to achieving safe agentic AI.

We contribute to open research and collaborate with teams to build robust, production-grade agentic AI systems.

About us

We're a world class team of researchers, engineers and prize-winning academics from Oxford University and Imperial College. We've built models competitive on global frontier benchmarks, and deployed agentic systems in sensitive environments ranging from FTSE 10 banks to cutting edge AI startups. We're based in London and are backed by leading technology investors.

Research

We're a small, highly collaborative team, and we love to work with academic experts to contribute to open research.

Currently, we have a particular interest in measuring healthy human-agent interactions, including detection of boundary violations in emotional behaviour. If you're interested in working on problems in this space, please reach out.

Latest writing

Radiant

Running agents in production? Radiant follows every job your agents run, finds what went wrong and tests the fix with simulated users. See getradiant.ai.

Press

If you are a member of the press and want to get in contact with us for information or comment please email [email protected]