I build the control layer around serious agents: runtime memory, tool routing, terminals, policy gates, trace capture, replay, and reviewable evidence.
The center of gravity right now is SpoonOS and the execution surfaces around it: making agent workflows observable, useful, and operator-grade.
| Project | What it proves | Surface |
|---|---|---|
| SpoonOS | Agent OS work across memory, tools, skills, and runtime surfaces. | Agents, MCP, runtime memory, Web3 AI systems |
| AgentProof | Signed flight recorder and policy firewall for AI agent runs. | Trace capture, policy guards, proof bundles |
| VibeShell | AI-native SSH, SFTP, tunnel, and local terminal workspace. | Tauri, Rust, React, local-first ops |
| Harness Architecture | Bilingual source-grounded notes on real agent harness design. | Astro, MDX, Codex, Claude Code, OpenClaw |
| build-your-own-agent | Loadable skill and scaffold for designing and diagnosing agent harnesses. | Python, skills, linting, diagnosis scripts |
| Lane | I tend to build |
|---|---|
| Agent infrastructure | Runtime traces, permission boundaries, replay tools, MCP workflows, verifier loops |
| AI-native operations | SSH/SFTP/tunnel tooling, local-first command rooms, observable automation |
| Agent OS surfaces | Memory, skill systems, tool routing, chain-aware automation |
| Documentation that ships | Bilingual docs, source trails, architecture maps, skill-based scaffolds |
| Hackathon execution | Demo-first products with real workflows, not wrapper demos |
Useful AI systems need three things before they deserve more authority: observable execution, scoped permissions, and replayable evidence.
That is the lane I like: small teams, fast ships, hard proof.