I build things to explore questions I can't leave alone. Sometimes that's software I use every day; sometimes it's research or a prototype that changes how I think. Then I follow the next question.

• ShallowSWE: a benchmark for whether coding agents finish routine tasks, and what each success costs.
• Alloy: typed AI functions that feel like ordinary Python.
• Errand: your usual commands, run on another machine, without keeping a remote checkout.
• agent: a coding-agent runtime built for programs, not people. One daemon, many named bots, JSON events out.
• SandFS: copy-on-write workspaces for agents. Read data lazily, fork an attempt without copying the source, and collect the files it changed.
• IceVault: a macOS app for automated archival backups.
• Atlas: a sketch of what an operating environment built for software agents would need.

Lately I've been thinking about how AI could help engineers make and verify decisions about the physical world.