Claude Code, Cowork, Codex, Antigravity, Manus, OpenClaw — every agent platform runs the same four ingredients underneath. Learn them once, using a personal assistant as the running example, and you can build in whichever one you land on.
An AI agent is software that can actually finish a task, not just describe how to do it. You give it a goal — "check my calendar and flag anything urgent" — and it plans the steps, uses real tools to carry them out, and checks its own work before handing anything back.
That's different from the chatbots most people have used. A chatbot answers what you type, one message at a time. An agent works toward a result and keeps going until it gets there — asking you first about anything risky, the way a good employee would.
Connects to the inbox, calendar, and notes you already use — nothing new to check, nothing to migrate.
Teach it once and it sticks, instead of re-explaining the same preference every session.
Every example in this guide works by typing plain English into a chat box.
The same four ingredients power Claude Code, Cowork, Codex, Antigravity, Manus, and OpenClaw alike.
Every agent repeats one loop until the result matches what you asked for — every platform in this guide is just a different coat of paint on it:
Strip away the branding, and that loop runs on four ingredients wired together.
The underlying model — Claude, GPT, Gemini. Swappable, and the ingredient people obsess over most while it matters least once the other three are right.
What turns a reply into a finished task — it keeps observing, thinking, and acting instead of stopping after one response.
What it's actually allowed to touch — your inbox, calendar, notes, a browser, a payment processor.
What it knows about you — your role, your business, your preferences — loaded in before it starts working.
The app that runs all four together is called an agent harness. Claude Code, Codex, Cowork, Antigravity, Manus, and OpenClaw are all just different harnesses running the identical loop — different cars, same engine. Learn to drive in one and the rest come free.
Easiest to pick up: a simple chat-style harness like Claude Cowork.
Most transparent about its own thinking: Claude Code — it shows you the loop as it runs.
Most autonomous, and hardest to set up: OpenClaw. Get everything below working reliably in an easier harness first, then migrate it once you trust it.
Onboard your agent the way you'd onboard a new hire — it can't do good work without knowing your business, your tools, or what it's allowed to do.
Decide what it's allowed to touch before you decide what to ask it. Read-only access is enough for everything in this guide — it doesn't need to send or delete anything on its own yet. If it's ever compromised or simply wrong, the worst case should be small.
Create one dedicated project or folder for this assistant, separate from anything else you use the harness for. Everything below — its instructions, memory, and tools — lives inside it.
Write a context file that's loaded before every single task — your role, your business, your working preferences. Every harness calls this file something slightly different, but it's always the same idea.
agents.md (Codex, OpenClaw) · CLAUDE.md (Claude Code) · GEMINI.md (Gemini) · "Custom instructions" or "Project instructions" in chat-style harnesses like Cowork. Same file, same job. Don't want to write it yourself? Ask the harness to interview you and draft it for you.
You are my personal assistant, not a chatbot. I'm [name], and I [what you do]. Your job is to reduce the number of things I have to remember, chase, or repeat.
Before MCP, an agent had to learn every tool's own language — custom code per app. MCP is the universal translator that sits in between, so any harness can speak to any tool the same simple way. Connect the three sources this assistant needs:
"Connectors," "Integrations," or "MCP servers" in your harness's settings. One sign-in or one server per tool — revoke access from the same menu any time.
No — once a tool is connected, the agent already knows what it can do; MCP exposes that automatically. Only mention a tool by name in your instructions when the choice between tools actually matters, like which of two calendars is the real one, or where expenses should get logged.
Chat models have memory built in automatically — every conversation quietly feeds a cloud memory you can't fully see or control. Agents don't. Create a second file, memory.md, sitting next to your context file from Step 3 — empty is fine to start. Whatever you correct today is forgotten tomorrow unless you set up memory on purpose — which is a feature, not a limitation: nothing bleeds in from unrelated work.
Kept apart from Step 3's file so fixed instructions and learned habits don't tangle — the agent rewrites memory.md itself as it learns, and reloads it every session. Keep both files well under 200 lines so rules stop contradicting each other.
With context, tools, and memory in place, simple prompts should already produce good results.
Ask it these, one at a time, and watch it reach for the right tool each time.
This is the payoff — the assistant crossing three sources instead of answering from just one.
Skills are standard operating procedures for your agent. Explain a process once and you never have to explain it again — and unlike memory, a skill won't clutter every other conversation, because it only loads when it's relevant.
There are two ways in. Use whichever fits the moment.
Plan it ahead: hand it source material — a course, a checklist, a past proposal — and say "build me a skill from this." Most harnesses ship with a skill-creator skill that packages it up for you.
Package it after: walk it through a process once, manually, refining as you go — then say "turn what we just did into a reusable skill." It writes the SOP from the conversation you just had.
Each skill gets its own file, usually literally called SKILL.md, in its own folder alongside anything it needs (checklists, templates, formulas). Easiest is to let the agent create it, the same way it created memory.md in Step 5 — but you can just as easily write one yourself:
--- name: meeting-prep description: Three-bullet briefing before a meeting, pulled from notes and recent email. --- 1. Look up any notes on the person by name. 2. Search email for threads with them from the last 30 days. 3. Summarize into exactly three bullets: who they are, what's outstanding, what to raise. 4. Never send anything — this skill only prepares a briefing.
Everything above was scoped to one assistant — the one that runs your mornings. The same structure repeats for whatever else you hand off next: a fitness assistant that reads your training calendar, a travel assistant watching your bookings, a household assistant tracking bills and renewals.
A skill, a context file, or an MCP connection can live globally — available to every assistant you build — or at the project level, scoped to just this one. Keep anything broadly useful (like a "make this shorter" skill) global; keep anything specific to one part of your life (like your travel assistant's frequent-flyer numbers) local, so it doesn't clutter context it has no business being in.
| Concept | Usually called | Does |
|---|---|---|
| Context file | agents.md · CLAUDE.md · GEMINI.md · "custom instructions" | Your instructions for the agent — loaded before every task |
| Memory file | memory.md · built-in memory | What it's learned — updated by the agent itself |
| Tool connection | MCP server · connector · integration | Lets it actually touch a tool, like Gmail or Calendar |
| Skill | SKILL.md · SOP · saved prompt | A packaged process it reaches for by name |