Build Your First AI Agent (Without the Hype)

"Agent" has been inflated into a buzzword, so let me deflate it. An agent is three things: a goal, a set of tools it can use, and a loop where it acts, sees the result, and decides what to do next. That is it. A chatbot answers one message at a time — you drive. An agent takes something like "triage my inbox every morning" and runs the whole sequence itself: read, decide, act, check, repeat. The difference isn't intelligence, it's whether you hold the steering wheel.

The mistake everyone makes is starting too big — "an agent that runs my whole team." That is Anthropic's own warning in Building effective agents: reach for the simplest thing that works, and only add agentic complexity when the task genuinely earns it. Start with the smallest agent that removes one real chore, then grow it.

The good news: you don't have to hand-build the loop from scratch. There are three real on-ramps, and they line up with how much plumbing you want to own. Start at the level that fits your comfort today.

Level 1 — a managed agent you never install (Claude Cowork)

The lowest-friction way in is Claude Cowork, Anthropic's background agent for knowledge work. There's nothing to install: in the Claude message box you pick Cowork, describe an outcome rather than a question, and it plans the steps, runs them on Anthropic's servers, and hands you the finished thing — a formatted doc, an organized folder, a synthesized report. Because it runs remotely, the work keeps going even if you close your laptop.

Give it a task with a clear finish line, not an open-ended wish. As an EM I use it for things like: "Go through the last 20 threads on this project; hand me a one-page status with the open risks and their owners." The trick is writing the goal as a brief, the same discipline as a good prompt, plus the one thing a prompt doesn't need — a definition of done:

Goal: a Monday digest of last week's on-call incidents.
Tools: read the incident channel export, write a markdown summary.
Done: dated digest, counts by severity, the 3 slowest to resolve, one trend line.
Constraints: under 200 words, no customer names.

• What makes it an agent: it breaks a request into subtasks and executes them end to end, not one turn at a time.
• Cost: bundled into your paid Claude subscription — Pro, Max, Team, or Enterprise — at no extra charge.
• Best for: anyone who wants time back today, with zero technical setup.

Level 2 — an agent that builds your tools (Claude Code)

Claude Code is Anthropic's coding agent, and it's the step where the agent stops producing documents and starts producing software. You describe what you want in plain language and it reads the codebase, edits files across the project, runs commands, fixes its own errors, and opens the pull request. You do not have to be able to write the code yourself — you have to be able to describe the outcome and check the result.

Install the CLI with curl -fsSL https://claude.ai/install.sh | bash, or skip local setup entirely and run it in the browser at claude.ai/code. Then ask for something concrete: "Build me a small dashboard that reads this CSV of leads, flags anyone we haven't followed up with in 7 days, and lets me export the overdue list." It plans, writes, runs, and iterates while you watch.

• What makes it an agent: it doesn't just draft code — it executes, debugs, commits, and can even run on a schedule or fan out into sub-agents.
• Cost: the tool is free to install; it runs on a paid Claude subscription.
• Best for: people who want real, custom tools — trackers, dashboards, internal scripts, automations they own.

Level 3 — a 24/7 agent on your own hardware (OpenClaw)

The top of the ladder is a fully autonomous agent that lives outside any single app and runs around the clock. The open-source project here is OpenClaw (MIT license, and one of the fastest-growing repos on GitHub). It runs a Gateway process on hardware you control and bridges it to the messaging apps you already use — WhatsApp, Telegram, Slack, Discord, iMessage and dozens more — so you talk to your agent the same way you text a colleague. It's model-agnostic: you bring your own API key from Anthropic, OpenAI, or Google and pick the brain.

What you need first:

• Hardware that never sleeps. Your laptop won't do — a 24/7 agent needs an always-on box. A base M4 Mac Mini (about $599, one-time) or a cheap VPS ($4–5/month, or an AWS free-tier instance) both work.
• Software. Node.js 24 (recommended) or 22.19+, then npm install -g openclaw@latest.
• An API key for whichever model you trust.

The setup runs about 30 minutes:

1. Install it — npm install -g openclaw@latest on your always-on machine.
2. Run the wizard — openclaw onboard --install-daemon. The daemon is what keeps it running 24/7 once the terminal is closed; the wizard then has you pick a model provider, drop in your API key, and configure the Gateway.
3. Connect a channel. Telegram is the quickest: message @BotFather, run /newbot, copy the token, and register it. For WhatsApp, run openclaw channels login and scan the QR code — use a dedicated number, and lock it down with an allow-list so only you can message it.
4. Add skills. Pull capabilities from OpenClaw's community skill registry — start with email, web browsing, and calendar so the agent can actually do things, not just chat.
5. Test it. Text your agent: "Summarize this link for me," or ask what your day tomorrow looks like. If it answers from your real tools, it's live.

• Cost: the software is free; you pay for hardware (a one-time Mac Mini or ~$5/month VPS) and API usage. A budget model runs a few dollars a month; a premium model at heavy volume can hit ~$140/month. Most personal setups land around $6–13/month total.
• Best for: ops-minded people who want genuine 24/7 automation across many tools at once.

Which level should you start with?

Start at Level 1. Build one Cowork agent that claws back 30 minutes of your day, and get comfortable handing over tasks and reviewing the output. Graduate to Claude Code once you need a tool that doesn't exist yet. Move to OpenClaw when you want work happening while you sleep. Each level teaches the muscle the next one needs.

From one run to running itself

The line between a clever assistant and a genuine agent is whether it runs without you starting it. Cowork crosses that line with scheduled tasks: type /schedule in any task (or open Scheduled in the sidebar), and it runs hourly, daily, weekly, on weekdays, or on demand — remotely, even when your computer is off. Claude Code has the same via routines, and OpenClaw runs on its daemon by default. Point it at a recurring chore — a Monday incident digest, a nightly build-failure summary — and it just shows up.

Anthropic's context-engineering guidance is blunt about what breaks first: most agent failures are context failures — the agent didn't have what it needed, or drowned in noise. So before you schedule anything, run it by hand a few times and fix the context, not the model.

This week

Name one chore with a clear done-state. Write the four-line brief above for it. Run it manually in Cowork three times and fix the context each time. By the third run you'll have something worth scheduling — and you'll understand agents from the inside, not the hype.

Sources

• Anthropic — Building effective agents
• Anthropic — Effective context engineering for AI agents
• Claude Cowork — Get started
• Claude Code — Overview
• OpenClaw — Getting started