Frontend and tooling engineer moving into AI tooling: the tools and harnesses agents use, and the product layer around them.

I'm learning this in the open. Coming from years of frontend, frontend build systems, and Node.js backends, I'm focused on what makes a tool actually usable by an agent: schema-first discovery, deterministic output, structured errors it can recover from. Sol is where I'm working it out.

What I'm building

• Sol the agent toolset for Upsun. The tool layer an agent or harness calls to operate the platform. Schema-first, token-efficient, built to be called, not read. heysol.dev

More agent-tooling work landing through 2026.

What I'm learning about agent tools

A tool an agent uses well tends to be:

• discoverable (schema-first)
• deterministic (stable, diffable output)
• recoverable (structured errors with hints)
• composable (one job per tool)
• non-blocking (no interactive prompts)

Where I'm coming from

Years across frontend, frontend build systems, and Node.js backends. The build-tooling side maps closely to agent tooling; the Node/TS side to MCP and SDK work; the frontend to the product layer. So the plan is to build AI products end to end, the tooling and the interface people actually use.