AI Infra Lead @ Yuanfudao (猿辅导 / Kanyun) · Beijing GitHub @krislavten · X @kris_wpf

🌏 中文版本 →

From observability engineer to AI engineer — building the infrastructure that makes AI-native software development possible.

I don't train models — I build agents. And not just the plumbing: I want to understand agents all the way down — what an agent really is, how you use it, what infrastructure it needs underneath. A heavy user of agents myself — 30B+ tokens a month.

Three things I'm exploring: agent coding — getting full-pipeline capabilities into production; agent applications — Rush for building websites, TentaClaw for observability; and agent infrastructure — turning what we've learned running agents in Rush into a managed-agents platform that speeds up the business and ships AI faster.

• agenthub (Pilot v2) — Self-built managed-agents platform: sandboxed Claude Code / Codex agents running coding (Git → PR/MR) and general (artifacts) tasks end-to-end. Four-layer resource model (Project / Workload / Revision / Session), credential Vault, resumable SSE, platform-level versioned Profile Registry — now serving 100% of production traffic. SDK: @kanyun-ai-infra/agenthub.
• Rush — Enterprise AI app/agent platform: describe what you want, get a full web app with live preview. 17,000+ projects, 3,000+ users, 500K+ messages internally; ~1,000 monthly active creators, 400+ agents. Core maintainer (512 PRs).
• OpenRush — Open-source, self-hosted managed-agents platform for Claude Code: sandboxed execution, Agent/Skill/MCP registry, dual-layer credential Vault, streaming event contract.

Tools so AI agents can operate real systems — all published to npm:

• octo-cli — Observability (logs / alerts / traces / metrics / RUM / topology)
• conso-cli — Console-platform ops (deploy / services / pods / resources)
• docz-cli / rush-shimo-cli — Company-docs access (DocSync / Shimo)
• jira-cli — Jira for agents (issue / sprint / board)
• qlint — Observability query linter — helps agents generate correct queries

• reskill (co-built) — Git-based skills package manager for AI agents — npx reskill install
• agent-aware — Lets agents perceive user behavior (clicks / scroll / frustration) in web apps
• Sparring — Adversarial AI code-review plugin: a second AI (Cursor / Codex / GLM / Claude) hunts bugs in Claude's output
• cowork-mdm — Claude Desktop enterprise deployment toolkit (reverse-engineered 51 MDM keys)
• rush-plugin · claude-env · env-sync — Workflow & dev-env tooling

• TentaClaw — Closing the loop beyond coding: Issue → Code → Deploy → Observe → Feedback, one agent per repo.

• I build the tools and live in them. Rush, agenthub and OpenRush are themselves built by AI agents (Claude Code + Cursor), with Sparring as the quality gate — a tight build/use feedback loop.
• Verifier-first; environment design over central orchestration. Define "how to verify it's correct" first, then let multiple agents self-coordinate through a shared environment instead of a central orchestrator.
• Pace: ~971 PRs / ~2,100 commits in H1 2026 (512 on Rush alone).

Before AI, 4 years building Octopus — Yuanfudao's company-wide observability platform (RUM / Logs / Trace / Alerts / LLM Observability). That's why the "Observe" loop is grounded in real production pain, not theory.

TypeScript · Node.js · React · Next.js · Hono · PostgreSQL · Redis · Docker · K8s Claude Agent SDK · AI SDK · MCP · Codex · OpenTelemetry · pnpm/turborepo

• Email — [email protected]
• X — @kris_wpf
• 小红书 — @kriiswu