Ethan Yan
Ethan Yan

About

A developer who gave the model hands.

I'm Ethan — an AI agent engineer who got curious about what happens when you give a language model hands. These days I spend most of my time on the unglamorous middle layer of agents: the retries, the guardrails, and the memory that keeps a system coherent past the first prompt.

Before AI I spent years shipping production systems — high-concurrency Java on the back end (banking-grade platforms handling tens of millions of transactions a day), then full-stack product work and a CS (AI) master's. That range is useful now: agents are systems, not prompts, and the hard parts live in the plumbing.

What I care about

The path

From backend to agents.

2026.06 Insight

Loop Engineering

You stop prompting the agent and start writing the loop that prompts it — guardrails are the real work.

2026.05

OpenFix

A local macOS agent that diagnoses and fixes Mac & network problems — no jargon, just answers.

AI Agent · macOS · Tool-use

2026.05

HR Agent

My current focus — a production HR/recruiting agent that writes code in a sandbox to call primitives, handling talent workflows end to end.

AI Agent · Code-action · HR

2026.05 Insight

Code-Action

Let the agent write code in a sandbox to call primitives, instead of piling on ever more tools.

2026.04

ReqBridge

Turns vague product requirements into structured, build-ready specs.

AI · FastAPI · React

2026.04

llm-eval-platform

Multi-model × multi-prompt evaluation with LLM-as-Judge scoring and resumable batch runs.

LLM · Eval · Next.js

2026.04

LLM Gateway + Observability

Self-hosted LiteLLM gateway with Langfuse tracing — one routing layer, unified keys, and full observability for every AI app.

LiteLLM · Langfuse · LLMOps

2026.04

Hermes vs OpenClaw

A hands-on comparison of two open agent frameworks — weighed capability, ergonomics, and cost, and landed on Hermes.

Agent frameworks · Evaluation

2026.04

OpenClaw Personal Assistant

Started building my own self-hosted assistant on OpenClaw — scheduled tasks, digests, and automations running on my own servers.

OpenClaw · Self-hosted · Automation

2026.04 Insight

Memory

Context isn't memory — what to keep, summarise, and forget between prompts is its own design problem.

2026.04 Insight

RAG

Retrieval to ground the model in real data — and that retrieval quality, not the model, is usually the bottleneck.

2026.03

claude-code-rev

Took Claude Code's source apart to understand how the agent harness works.

Agent harness · TypeScript

2026.03

Smart Building Report

An agent system that turns building-inspection field data into structured, ready-to-send reports.

AI Agent · Reports · Python

2026.03 Insight

Agent Harness

Pulling Claude Code apart made the 'harness' click — what equips a single agent run.

2026.02

pingcode-mcp

First MCP server — standardised, pluggable tool access for agents.

MCP · TypeScript

2026.02 Insight

Context Engineering

Context is a finite attention budget — the job is curating what goes in.

2026.02 Insight

LangChain / LangGraph

Frameworks made agents buildable; LangGraph's explicit state and graphs mattered more than the chains.

2026.02 Insight

MCP

Tools as a standardised, pluggable interface rather than bespoke glue.

2026.02 Insight

Skills

Packaging capability as modular, on-demand skills — the agent loads what a task needs instead of carrying everything at once.

2025.09

SaveVia

Canadian credit-card cashback AI advisor, shipped to the App Store — with my first real ReAct agent inside.

FinTech · ReAct Agent · React · Spring Cloud

2025.09 Insight

ReAct Agent

First time letting a model pick its own tools and loop on the result (async ReAct, 5-round cap).

2024.07

VAMerchant

Production multi-tenant merchant SaaS — appointments, scheduling, billing, multi-language.

SaaS · Microservices · AWS

2016

Backend foundations

Seven years of backend engineering before AI — banking-grade high-concurrency systems (30M+ transactions/day, 99.99% uptime) and a CS (AI) master's.

Java · Distributed systems · MSc CS (AI)

Get in touch

Email me at yancheng0723@gmail.com, or find me on GitHub. Always happy to talk about agents and AI engineering.