Vision · June 2026
You supervise, AI executes
The next decade of software is humans supervising agents — and the interfaces, protocols, and trust models all have to be rebuilt around that premise.
Every product we build comes back to one sentence: you supervise, AI executes. This is the longer version of why we think that's the right bet — and what it demands from the software we build.
The premise
For most of computing history, humans operated software. We clicked, typed, configured, and the machine did exactly what it was told. The interesting shift of the agent era is that this inverts: increasingly, the machine acts and the human judges. You don't operate the agent step by step — you give it intent, it executes, and your job becomes supervising the result.
That sounds like a small reframing. It isn't. Almost every interface, protocol, and trust model we have was designed for the operating model, not the supervising one.
What supervision actually costs
The scarce resource in a supervised system isn't compute — it's human attention. So the central design metric we care about is interventions per completed task: how often does a human have to step in to keep the agent on track?
Drive that number down and the system scales with the person. Leave it high and you've just built a more elaborate way to do the same work. Most "AI features" quietly fail this test: they generate output fast but demand just as much vigilance to verify, so the human bottleneck never moves.
A good agent doesn't just do the task. It makes itself cheap to supervise — easy to check, easy to correct, honest about what it's unsure of.
Three things have to be rebuilt
If supervision is the job, then:
- Interfaces have to surface the right thing to approve at the right moment — not bury the decision in a wall of output.
- Protocols have to let agents coordinate with each other and report back in a form a human can audit — agents negotiating directly, with the human reading the outcome, not the transcript.
- Trust models have to be earned incrementally — the agent does more unattended only as it proves it's cheap to be wrong.
How we work
We ship in public, one product at a time, and write up what we learn. The research notes here track the outside work shaping our thinking; this blog is where we think out loud about the product side.
The thesis won't be right in every detail. But the direction — humans supervising, AI executing — is the one we're building toward, and everything on this site is an attempt to make that loop a little cheaper to run.