May 23, 2026

Building AI Agents That Actually Work

Andrew Luo Weimin

345 Words 1 Minute, 34 Seconds

2026-05-23 11:31 +0000

1. Chat vs. Agents

Chat Models: A simple ping-pong interaction (Question $\rightarrow$ Answer / Back-and-forth).
AI Agents: A goal-driven execution model (Goal $\rightarrow$ Result). You specify the outcome, and the agent plans and executes autonomously to achieve it.

Agents operate in a continuous loop:

Observe: Research, gather context, and analyze the task environment.
Think: Plan the workflow, formulate steps, and evaluate options.
Act: Execute actions (write code, run commands, fetch resources) and verify the outputs.

An effective agent setup is treated like an Operating System, structured using specific configuration files and folders:

agents.md: The onboarding document for the AI. It defines the business voice, user context, expectations, and the agent’s role.
memory.md: A living file updated dynamically. It stores user preferences, rules of thumb, and learned behaviors (e.g., specific formatting rules, things to avoid) so the agent self-improves and becomes more personalized over time.

Context Folder: Contains rich reference documents, templates, and deep background material relevant to the project or business.
Skills Folder: Reusable Standard Operating Procedures (SOPs) for the AI. Once a complex workflow is solved, package it into a “skill” (a structured prompt/script) so the agent can execute it autonomously in the future.

Agent Harnesses: Platforms like Claude Code, Antigravity, Codex, Cowork, and Manus act as the “chassis/car,” while the LLM acts as the “engine.” Learning how the agent loop works in one tool allows you to operate any of them.
MCP (Model Context Protocol): An open standard created by Anthropic that acts as a translator, allowing LLMs to safely connect to external tools (Gmail, Notion, Slack, databases, browsers) without custom development.

Automate SOPs: Shift from manual administrative work to building automated skills. Once a skill is written, that task is automated forever.
The Math of Compounding: Automating 3–5 small tasks a week builds massive leverage, eventually allowing individuals to run entire automated departments and compress 40 hours of work into a single afternoon.