The lab
Half of this job is holding on to the questions you can’t answer yet. This is where I keep mine: open threads, half-formed opinions, and a few things I’ve changed my mind about.
Where agent autonomy actually saves time, and where it just adds risk.
Experimenting with autonomous AI agents, custom skills, and multi-step flows. Currently exploring what an agentically run brand experience could look like beyond the conventional website.
How multi-agent systems negotiate, delegate, and recover from failure.
No conclusions attached. I’m leaving them as questions on purpose. There’s a surprising amount going on under the surface of the everyday world.
The Developer Knowledge MCP server didn't ship because a demo looked convincing. We tested it with a blind evaluation: a technical writer scored answers without knowing which version produced them. The score improved from 7.6 to 8.5, surpassing the launch bar we'd set at 8. Good demos show what's possible. Rigorous evaluation tells you what's ready to ship.
The eval storyFirebase Studio's Chat Unification charter arrived as an open-ended instruction: unify the experience. It only became buildable once it was scoped into two actual questions, unification then context sharing, each with its own milestone. Ambiguity usually isn't a property of the problem. It's a sign the question hasn't been cut in the right place yet.
How it was scopedThe interesting part is how different pieces work together: what an agent should retrieve, which tools it should invoke, what it should remember, and how it should decide what to do next. Less about making a model know more, more about giving it the right capabilities, context, and autonomy to get things done through orchestrated agent workflows.
The shift is from AI that answers questions to systems that can reason, act, and deliver outcomes.
How it's built