Skip to main content
EMil Wu
中

AI Agent Article Series

AI Agent Developer Notes

From Model ≠ Runtime to Workflow Mindset — 11 in-depth articles covering the technical foundations and mental model shifts of AI Agents. If you have a technical background, start with the 7 Technical Framework articles, then read the 4 Mindset articles. If you're less technical but want to learn AI workflows, start with the Mindset articles and refer back to the Technical Framework when questions arise. I hope these articles help — if you find them useful, consider buying me a coffee using the button at the bottom of the page~

Start Reading

News

Latest AI Agent Updates

More →

AI Traffic Is Testing GitHub, and Even Mitchell Hashimoto Is Leaving When Actions Stalls

Ghostty leaving GitHub is not just a sentimental farewell from a long-time open source maintainer. Read alongside GitHub's recent Actions delays, its own 30X capacity planning language, and the surge in agentic development workflows, it points to a new pressure point: AI automation is starting to stress the PR pipeline developers have treated as background infrastructure.

Eight Cents an Hour: Anthropic Will Run Your Agents for You

Anthropic launches Claude Managed Agents in public beta — define your agent logic, and they handle sandboxing, state management, error recovery, and security. At $0.08/hr with an OS-inspired 'brain-hand decoupling' architecture, plus Notion and Rakuten deploying in under a week, is the Agent-as-a-Service era officially here?

Milla Jovovich's AI Memory Tool: The Perfect Score That Didn't Survive the Week

Actress Milla Jovovich and engineer Ben Sigman built MemPalace, an open-source AI memory system, using Claude. It claimed a 100% perfect score on the LongMemEval benchmark and racked up 5,400 GitHub stars in 24 hours. Days later, developers found the scores were measuring the wrong thing entirely.

Practical Tips

Real-world Collaboration Insights

Practical Tips 4: The Agent's Timezone Blind Spot — A 25-Hour Gap It Never Noticed

A daily email assistant Agent accumulated a 25-hour data gap due to timezone assumptions. Worse, even with evidence right in front of it, it never noticed it was missing emails.

Practical 8 min

Practical Tips 3: The Agent Reading Trap — Summaries Speed You Up, and Make You Miss

Agents default to summarized reading for everything, including code. Extracting Claude Code's 3,200-line /insights source revealed: three reading depths produce three quality levels — because subtle differences in code aren't redundant, they're deliberate design decisions.

Practical 8 min

Practical Tips 2: Packaging Your Skills as Plugins for Distribution

A real-world walkthrough of the Claude Code Plugin packaging workflow: structure design, dependency classification, environment variable handling, verification loops, and building a unified Marketplace. Includes a full pitfall log.

Practical 10 min

Practical Tips 1: The Agent's 'Known' Trap — What It Doesn't Say Is the Most Dangerous

Agents treat session context as implicit knowledge — omitting key info, hiding capability gaps, and turning temporary constraints into permanent rules. Three real-world cases reveal 'input-side hallucination' and a daily handoff ritual to prevent it.

Practical 12 min

Agent Team in Practice

Building a collaborative Agent team from scratch

Agent Team in Practice (11): Persistence Is Not Keeping Agents Running Forever, but Letting Them Wake Up Ready to Work

From scheduling a remote agent to write news overnight to A7sus4's persistent identity, this article organizes long-run persistence for Agent Teams: how memory, task, identity, and observation persistence connect so agents remain handoffable, traceable, and verifiable across sessions, machines, and time.

Agent Team 7 min

Agent Team in Practice (10): A7sus4 — I Did Not Build A7 Directly

After A7's role became clear, I did not build A7 directly. I built A7sus4 first: a production precursor that carries Role 3 production duty. This article covers persistent identity, signal-density preflight, the first live audit findings, and the methodology for building auditing Agents that grew out of A7.

Agent Team 6 min

Agent Team in Practice (9): A7 — Not One More Agent, but the Team Learning to See Itself

A7 began as a future answer to a validation gap, then grew through silent failures, chain debug ownership, and team drift into a real Auditing Agent. This article focuses on why A7 needs to exist: not as one more executor, but as the role that lets the Agent Team see its own failures.

Agent Team 6 min

Agent Team in Practice (8): Communication Architecture — What OS IPC Can and Cannot Teach You

OS assumes all processes are homogeneous, but every Agent in an Agent Team is heterogeneous — this fundamental difference makes OS IPC fully applicable at L0-L2 (Environment, Transport, Topology) but breaks down at L3-L4 (Protocol, Content Contract). Five problems the OS framework can't solve: shell hooks can't extract LLM judgments, the filesystem is both transport and result, unified formats are meaningless across heterogeneous Agents, receiver Context degrades with excessive tokens, and identity loading is a dimension OS never needed — each grew its own solution: inspection maps, Return Contracts, and Hybrid Dispatch that no OS textbook covers.

Agent Team 8 min

Workflow Mindset

Articles #08–#11

Unprofessional Arrogance: When AI's Capability Becomes Your Illusion

Non-engineers overestimate their cognitive abilities thanks to AI — unprofessional arrogance. Reverse Dunning-Kruger, 17x error trap, Replit database deletion — AI works 99% of the time, but the fatal 1% hides where you can't see. The narrowest path between two arrogances is called humility.

Mindset 3 min

Professional Arrogance: When Engineering Experience Becomes AI's Ceiling

Engineers distrust AI due to experience-based frameworks — professional arrogance. METR research shows senior developers' productivity dropped 19% with AI — not because AI failed, but because they spent too much time questioning it. Expert + AI is still the best combo, but only if experience doesn't become a wall.

Mindset 4 min

Golden Circle 4: From Playbook to Execution — Plans Change, That's the Point

A Playbook is a recipe; an Execution Plan is tonight's prep list. From contextualized decision-making to a real case where the v1 plan hit a wall, understand why plans changing is normal — and what it means for Agent collaboration.

Mindset 8 min

Golden Circle 3: From Methodology to Playbook — The Three Operationalization Gaps

Methodology can't directly become a Playbook. The operationalization process requires three layers: facts, scenarios, and tools. From a 490-line mixed document split into three separate files, understand the essential difference between methodology and playbook.

Mindset 6 min

Technical Foundations

Articles #01–#07

Agent Team: When Subagents Aren't Enough

When a task requires multiple Agents to collaborate rather than simply being dispatched, Agent Team breaks the hub-and-spoke ceiling using mailbox communication channels and a shared task list — but the cost is 15x the token consumption.

Technical 5 min

Skill + Subagent: Combination Patterns and a Decision Model

Skill manages knowledge loading; Subagent manages Context isolation. The two can be combined into an Explore → Decide → Execute three-stage workflow, and cost considerations make the architecture decisions even sharper.

Technical 4 min

What's Actually the Difference Between a Skill and a Subagent?

Skills and Subagents don't live on the same layer. A Skill is the main Agent's knowledge module; a Subagent is an independent executor with its own Context. Choosing between them comes down to one question: does the intermediate process belong in the main conversation?

Technical 6 min

The Next AI Battle: Model Supremacy or Skill Ecosystem?

Putting the first three articles together reveals a clear picture: the center of gravity in AI competition is shifting away from the Model layer and toward the Skill ecosystem. Whoever builds the richest Skill ecosystem first holds the strategic high ground in the next round of competition.

Technical 4 min

Peripheral Resources

Tips & References

Running Multiple Claude Code Accounts at the Same Time

Three approaches for using multiple Claude accounts simultaneously, from simple to advanced.

Commands → Skills Migration Guide

How to migrate .claude/commands/ to .claude/skills/, with a downloadable Agent Playbook.

Extracting Claude Code /insights from the Leaked Source — Install & Usage Guide

Reverse-engineering the /insights command from Claude Code's leaked source, repackaged as a Skill and Command with install instructions.

Support This Series

If these articles have been helpful, consider buying me a coffee