I care about where software meets judgment — the thin layer where agents, products, and humans actually collaborate.

Most of my work lives in three currents:

Local-first agent runtimes, multi-agent collaboration, tool-use loops, and the infrastructure that makes long-running AI work trustworthy.

Native desktop utilities, browser feedback channels, and UX that turns vague intent into precise machine action — without a wall of prompts.

Full-stack delivery across TypeScript and Go: monorepos, BFF/gateway patterns, observability, and shipping under real production constraints.

Day job is large-scale product engineering at ByteDance. Public writing and open source stay intentionally abstract — the interesting details are usually the ones that shouldn't leave the building.

From recent local craft — without naming internal products:

• Coding agents & workbenches — orchestration, skills, eval harnesses, and desktop shells around Claude Code / Codex-class runtimes
• Visual feedback for agents — browser-side annotation that turns UI intent into structured agent input
• Native utilities — macOS tools that respect the OS instead of fighting it
• Small games & engines — canvas physics, browser 3D, games as a lab for deterministic simulation
• Language surface area — TypeScript for product velocity, Go for durable services

• Designing agent collaboration that feels like a team, not a chat box
• Collecting sharp tools at the edge of AI × systems × craft
• Open to thoughtful collaboration on open-source agent infrastructure and developer experience