Build the platform that turns observation into a working agent: the discovery that learns how a process really runs, and the runtime that executes what it produces.
Build the tools both sides call - into ERPs, inboxes, documents and screens - so a generated agent can touch the messy real world.
Make generated agents trustworthy: evals, shadow-mode comparison against real historical work, and the guardrails that decide when a human steps in.
Make the next process cheaper than the last. The platform is the leverage - never any single agent.
What we look for
Red-pilled on AI: you think the days of writing code by hand are done, and you've already stopped writing it that way.
You've built harnesses - the scaffolding, tools and context that let an agent actually finish a job rather than demo one.
You've run loops, and you've been down in the agent slop: the drift, the confident nonsense, the runs that go nowhere. You know how to get real work out the other side.
Rigour about accuracy: you measure before you claim.