Agent architectures and execution
How should an AI agent break down tasks, call tools, and handle errors without derailing?
Research for AI that works in the real world
Capable models are not enough. We research how to make AI systems reliable, controllable, and safe once they meet real data, tools, and people.
“A safe AI system is one where access, actions, and oversight are designed in from the start, not patched in later.”
Focus areas
Our research sits between powerful models and real deployment, where most AI systems succeed or fail.
Systems
How to build the software around a model so it can work with real data, tools, and workflows.
Safety
How to limit what an AI system can do, catch failures before they cause harm, and keep people in control.
Safe by construction
Designing access limits, permission boundaries, and audit logs into the architecture from the beginning.
Failure detection and handling
How can a system recognize when an output is unreliable or dangerous before it is acted on?
Human oversight interfaces
When and how should a human be brought into the loop, and what information do they need to make a good decision?
Evaluation
How to measure whether an AI system is reliable on real tasks, not just on standard benchmarks.
Task-based evaluation
Testing AI systems on the actual work they are meant to do, with realistic constraints and imperfect inputs.
Failure mode analysis
Systematically mapping the ways systems break in practice, and designing tests that catch those specific failures.
Continuous monitoring
How do we know a system is still behaving correctly after deployment, as data and conditions change?
Where we're starting
Our early work is focused on foundations: how AI systems should be built, constrained, and tested.
View research
Featured roles
We are an early team looking for people who want to work on hard, open problems in AI systems.
View all careers