Product leader and builder. I make agent-first software.
I'm interested in one question: how do you make AI agents do real work reliably? My answer so far is deterministic orchestration, context that lives in git, and verification that is cheap enough to automate.
- hivekit — Terraform for software. You write a
HIVE.mdthat says what you want and how "done" is verified. A Planner → Generator → Evaluator loop builds and reconciles until the criteria hold. Apache-2.0, built on the Claude Agent SDK.
- Intent is the durable asset. Models and agents are disposable. The file that says what "done" means is not.
- Context is truth. My personal operating system is a git repo that agents read and write.
- Autonomy is bottlenecked on verification, not on loops. Where the done-check is cheap, automate hard. Where it needs human judgment, keep the gate explicit.