The arc from
intelligence to action.
We build autonomous agents that operate real systems — not chat about them. Progress here is counted in tasks completed, never benchmarks won.
Step 01 — Observe
Reliability, not capability,
is the wall.
Every frontier model can describe almost any task. Very few can complete one end to end, inside a real system, without a human catching the failure.
deployment quietly dies. We only work on the second bar.
Step 02 — Plan
Six problems between a model
and something deployable.
None of these are solved by a larger model. All of them are engineering and evaluation problems, which is why a lab has to work on them directly.
Grounded execution
Acting inside real environments — terminals, APIs, browsers, internal tooling — with verified state after every step, not assumed state.
Recovery, not retries
Detecting failure mid-task and re-planning around it, instead of repeating the step that just broke and hoping.
Completion metrics
Evaluating on whether the task actually finished in the real system. Benchmark scores do not transfer to production.
Bounded autonomy
Explicit permissions, reversible actions, a knowable blast radius. An agent should never be able to exceed what you sanctioned.
Legible traces
Every decision inspectable afterwards, in plain language, by someone who wasn't watching it happen.
Long-horizon work
Tasks measured in hours and days, holding intent across sessions without drifting off the objective.
Step 03 — Execute
Everything is measured
on the last step.
A run is not partially successful. Either the invoice posted, the ticket closed, the migration finished — or a human is now cleaning up after your agent.
Reproducible or it didn't happen
Every claim we publish comes with a trace we can re-run on request.
End-to-end only
The single unit of progress we recognise internally. Partial credit isn't credit.
No demo we can't run live
If it only works in a recording, we don't show it at all.
Step 04 — Fault detected
Per-step accuracy is a comforting number. It is not the number that matters.
An agent that needs supervision to be trusted has not automated anything —
it has moved the work from doing to watching.
You just watched this page fail. Keep scrolling — that part matters more.
Step 05 — Recover
Recovery is the product.
Anything can succeed on a clean path. The lab's actual work begins at the moment something breaks: notice it, understand why, re-plan around it, and still finish — without a human in the loop.
That's the difference between a model that is capable and an agent you can deploy. It's the reason we measure completion instead of accuracy, and it's the only benchmark we care about improving.
Step 06 — Log
Working in the open.
Short, specific write-ups on what we're learning. No press releases, no launch posts.
The 95% trap
Why per-step accuracy collapses across multi-step tasks, and what to measure instead.
Failure is information
Treating a broken step as a signal to re-plan rather than an error to repeat.
Benchmarks don't transfer
The measured gap between evaluation scores and completion inside live systems.
Knowable blast radius
Designing permission surfaces so an agent's worst case is understood before it runs.
Step 07 — Handoff
Give us something
that actually breaks.
Early partners
We're taking a small number of partners with real operational workloads — the messier and more failure-prone, the more useful to us. Researchers and engineers who want in, say so directly.