I build autonomous AI agents — systems that can actually run on their own. Not wrappers around a chat interface, not prompt-and-response loops. Things that observe their environment, make decisions, take action, and course-correct without me in the loop.

That's the part I find genuinely interesting. Getting an agent to do something once is easy. Getting it to do something reliably, recover when it fails, and know the difference between the two — that's the actual problem.

Right now I'm focused on the full agent lifecycle. Giving a system real context about where it is and what it has access to. Building the reasoning layer that turns that context into a plan. Then the execution layer that carries the plan out safely — with validation, retry logic, and feedback so the agent knows whether it actually worked.

I'm also into local-first AI. There's something that just clicks about a system that doesn't depend on an internet connection to function. Self-contained, reproducible, yours.

Context is everything. A model with no awareness of its environment is just autocomplete with good PR. What makes an agent useful is knowing where it is, what it can touch, and what happens if it gets something wrong — before it acts.

So I care a lot about the layer between LLM output and real-world execution. The boring infrastructure that makes the interesting behavior possible. Sandboxing, validation, feedback loops. That stuff isn't glamorous but it's what separates a demo from something that actually runs.

I try not to be precious about tools either. Whatever solves the problem cleanest is what I use. The stack follows the work, not the other way around.