Lee's Notes
Building AI-native tools, coding agents and scalable systems.
Latest Posts
-
AGENTS.md Is a Constitution, Not a Brain: Layering Context for AI Coding
A few hundred lines of AGENTS.md and the agent still misses required interfaces, skips confirmations, and treats stale rules as current. This post splits a coding agent's environment into bounded layers — AGENTS.md, Spec, Wiki, Skill, CLI/MCP, and Review/CI — and shows what each layer should (and shouldn't) be responsible for.
-
Claude Code's Most Powerful Tools Write No Code
After dissecting the tool definitions of Claude Code, Pi, Codex, DSH and other coding agents, I found that what makes Claude Code special isn't what it can do — it's what it keeps itself from doing. PlanMode changes no code yet is the pivot of its whole controllability architecture; Edit's real value is stopping the model from editing recklessly; Bash can do anything, so it's fenced in at every layer.
-
How Much Should an Agent's Prompt Carry? — Pi vs Claude Code on Where Harness Complexity Lives
154 lines vs 1627 lines of System Prompt: Neither Pi nor Claude Code made complexity disappear — they just moved it somewhere else.
-
DeepSeek’s Agent Framework: Even the Agent Loop Is a Plugin
A deep dive into DeepSeek Harness — 40+ packages, 20+ capability seams, and an Agent Loop that can be swapped from a YAML config. Capability Seam three-role separation, the Session Log as the single source of truth, and a 'model-visible ⟺ logged' runtime invariant.
-
TencentDB Agent Memory: Share Experience, Not Privacy
A deep dive into Tencent's open-source Agent Memory project — layered memory architecture, four asset types, Loadout mechanism, and comparisons with Mem0, LangGraph Memory, and Zep.
-
How to Report at Work: A High-ROI Skill
A practical guide to reporting at work — the reality of the information gap, four common traps, one framework that is enough, and how to make data tell a business story.
-
Pi Agent Design Philosophy: From Bare-Bones to Light Harness
What Pi Agent cut away tells you more than what it kept. Six key decisions behind its "light harness" philosophy, plus 5 takeaways for building your own harness.
-
All in Agent: Integrating Workflows into the Agent
In the Agent era, the default entry point for operations is no longer the platform UI, but the Agent itself.
-
Agent Memory Management: A Layered Framework for Long-Term Context
A practical framework for organizing what agents should remember, what they should forget, and when to update — answering what to keep, what to discard, and when to promote.