The through-line
The interesting frontier is not "an agent can call a tool." The interesting frontier is how to make tool use inspectable, recoverable, safe, and useful inside real workflows.
That shows up in slash-command UX, skills, reusable prompts, agent teams, MCP tools, secure sandbox installation, human approval boundaries, and operational loops that make the agent useful after the demo ends.
What I keep testing
The public experiments are small but pointed: command layers that make agent capabilities discoverable, reusable skill packs, tool contracts, sandbox setup, agent teams, and workflow-specific kits such as meta-ads-kit where human approval and tool execution need to meet cleanly.
What I keep learning
This project family is the public bridge between my platform work and my personal build habit. It shows interest in the whole stack: local developer workflow, agent runtime ergonomics, sandboxing, tool contracts, and the weird human factors that determine whether AI tooling is actually adopted.