The Agent Landscape

Today you installed Codex and directed your first build. Codex is not the only agent, and this course will not pretend it is. Over four weeks you'll work with two of them and learn the patterns behind all of them — because the frameworks change every year, and the person who understands the landscape doesn't have to relearn it each time.

The shape they all share

Strip away the logos and every agentic framework is the same four parts:

A model — the intelligence that reads and decides (GPT-5.5, Claude, and others). Tools — the hands: creating files, running programs, browsing, sending messages. A workspace — the folder or environment where the work happens. Memory — what survives between sessions: notes, instruction files, learned preferences.

When the agent works, these parts turn in a cycle: the model reads the situation, picks a tool, acts, looks at the result, and decides the next move — around and around until the job is done. That cycle running unattended is what people mean by an agentic loop. You'll build your own loops in week 2; today, just recognize the shape.

Codex — the builder you already met

Codex is OpenAI's agent for making software. You've seen the desktop app: threads organized by project, a chosen working folder, changes you can inspect before accepting. Under the hood it runs in a sandbox — it acts inside your chosen project, not across your whole machine. It also exists as a command-line tool and inside code editors; the app is simply the friendliest door.

Its economics matter for this course: Codex runs on the ChatGPT subscription you bought today, metered in usage windows rather than per-token bills. For the next four weeks, Codex is your daily instrument.

Claude Code — the specialist's coding agent

Anthropic's Claude Code answers a narrower question with unusual depth: how good can an agent be at serious software work? It lives in the terminal and in editors, reads large codebases well, plans before it acts, and has a rich ecosystem of skills — reusable instruction packages the community writes and shares. Professional engineers often describe it as the strongest pure coding agent; its trade-off is focus. It manages code, not your calendar.

In this course Claude Code appears in demonstrations — including a fact worth knowing: the course you are reading was designed with Claude and built by agents directing each other. When we reach orchestration in week 4, that example becomes a lesson.

Hermes — the agent that learns you

Hermes (by Nous Research) starts from a different premise: an agent should get better the longer you use it. It keeps its knowledge in plain-text memory files — readable documents about your projects, your preferences, your corrections — and it revises those files itself as you work together. Correct it once and the correction is written down; next week's session starts smarter than this week's.

That evolving memory also gives you more control over what the agent keeps in mind — which becomes important when projects grow. Practical detail: Hermes signs in through the same ChatGPT account you created today, so it costs nothing extra. You'll install it in week 2, and that's the day we study memory files properly: what agents remember, where it lives, and why it changes everything about long projects.

OpenClaw — the agent that lives on a server

OpenClaw is the largest open-source agent runtime — hundreds of thousands of GitHub stars, millions of users. Its philosophy: your agent shouldn't live in an app you open; it should live where you already are. OpenClaw runs continuously on a small server, connects to Telegram, WhatsApp, Discord and more, runs scheduled jobs while you sleep, and routes work to whichever AI model you give it.

Message it “remind the team about invoices every Friday” from your phone, and there is no app to open — the agent is simply present in your messenger. In weeks 3 and 4, when this course turns to phone-controlled agents and customer-support bots, OpenClaw's patterns are the ones you'll be using.

Where Pi fits — and doesn't

You may also meet Pi, a warm, conversational AI companion. Pi is pleasant to talk to by design — but notice what's missing: no workspace, no tools, no files that exist after the conversation. Pi is chat. It's a useful contrast to keep: a good conversation is not delegation. If nothing exists on a computer afterwards, no agentic work happened.

Side by side

Check yourself

You don't need to choose a side. You need the map: builders at the desk, runners on the server, and memory that either resets or learns. Tomorrow you go back to the desk — there's a game to build.