About
The oldest problem in automation
Heron Lab builds AI agents for the real world. We work on the ones whose tasks carry consequences, and we hold them to the standard of instruments rather than demonstrations.
Namesake
The name comes from Heron of Alexandria, who wrote a treatise on automata in the first century and built a mechanical theatre that performed an entire play by itself. The program was a cord wound around an axle. Change the winding and you change the play.
It worked because the stage never surprised it. Ours does, constantly: a source goes quiet, a step that worked yesterday fails for a reason nobody wrote down. An agent that only holds when nothing moves is a theatre piece, not an instrument.
01
Ground truth first
An agent is only as good as its observation of the world. We invest in instruments before intelligence: verification, confirmation, and fresh state at every decision.
02
Institutions, not prompts
Complex work is organized work. Our systems have roles, protocols, ledgers, and recorded dissent, because a team that keeps no minutes repeats its mistakes with complete confidence.
03
Consequences as curriculum
Deployment is where learning about consequences begins. Every confirmed failure becomes a regression test; the corpus of things that went wrong is our most valuable dataset.