Posts
All the articles I've posted.
-
In Agent Land, Errors Are Scheduled Work: From Error Boundary to Verifier
The model confidently declares "done" while the work isn't — an error that doesn't error, and no layer of try/catch will catch it. Mechanical failures get retry with exponential backoff plus a hard max_steps brake; semantic failures get an independent verifier — and verification must run in an isolated context. Part 6 of the fe2agent series, with tiny-agent v0.6 (withRetry + MAX_STEPS + verifier).
-
Context Is RAM, Memory Is Disk: Give Your Agent Lazy Loading
You think the agent remembered your rule; really that sentence lives in yesterday's session's messages array, and today it wakes up brand new. Context is RAM, memory is disk — "write it down" is the only persistence, and retrieve-and-backfill is agent-land's lazy loading. Part 5 of the fe2agent series, with tiny-agent v0.5 (a markdown memory file + naive keyword retrieval + two new tools, save_memory / search_memory).
-
Prompt Is the New CSS: Declarative, Cascading, No Devtools
Change one line of the system prompt and the agent behaves like a different animal — prompts are like CSS: declarative, cascading, overridable, but with no devtools to show which rule is in effect. Part 4 of the fe2agent series — prompt layers, instruction dilution, and structured prompting, with tiny-agent v0.4 (same tools, two system prompts, run the behavior diff).
-
From OpenClaw to Claude Code: A Half-Year Config Evolution Ledger
Two TODOs and one regret from three months ago are all settled, plus four things I never imagined: memory as the rule hub, Skills in the pipeline, bilingual i18n as a second publishing line, and AdSense reshaping the writing discipline. Seven evolutions, one section each — a ledger.
-
Tool Call: Let the AI Dispatch, You Stay the Reducer
In frontend you dispatch and you write the reducer. In an agent the AI dispatches; the reducer half is still yours. Part 3 of the fe2agent series — the tool schema trio, separation of execution rights, validating the AI like an untrusted client, with tiny-agent v0.3.
-
The Model Has No Memory: Context Is State, and Every Turn Is a Full Re-render
You think the agent remembers you; really it re-sends the full messages array every turn. LLM calls are stateless — "conversation" is a client illusion.
-
An Agent Is Just a While Loop? From UI = f(state) to action = LLM(context)
An agent is a while loop whose function isn't pure. Series opener from UI=f(state) to action=LLM(context), with a 32-line runnable tiny-agent.
-
3 Principles Fiber Teaches Us About AI Agent State Design
How Fiber's three mechanisms (double buffering, time slicing, lanes priority) map to AI Agent state design — three principles I stole and dogfooded.