01Applied AI● Active
Better evals for AI agents
Multi-step AI agents are hard to evaluate. We're building a way to measure whether agents make good decisions — not just whether each step is correct.
64%PROTOTYPE
Research
What we're working on.
Half of our work is for clients. The other half is research — long-horizon bets that make the client work better.
Active projects
In progress.
02Infrastructure● Active
Replay-able distributed systems
Generalising what we built for Project Lattice. The aim: systems where you can replay yesterday's state exactly, by design — not by accident.
41%RESEARCH
03Interfaces● Early
Ambient computing surfaces
Software you notice at a glance, not at a click. Small physical prototypes — a lamp, a wall display — that share one design language.
22%EARLY
04Cross-team● Active
Synthetic test data from real traces
Most test suites cover the easy cases. We're generating adversarial test data from real production traces — open-sourcing the harness this quarter.
78%SHIPPING
Future trajectories
Where we're headed.
2026 – 2027
Autonomous code review
Agents that review pull requests the way a senior engineer would — checking for correctness, not just style.
2027 – 2028
Zero-config observability
Instrument any codebase in under a minute. No YAML. No SDKs. Just understanding.
2028+
Reasoning over long contexts
Models that stay coherent across very long documents and multi-day task sequences. The infrastructure problem no one has solved.