You can build an AI agent with Claude in three setups; the first is a preliminary stage, only ways 2 and 3 work agentically: as a project in Claude.ai (no coding, inside the app), as a subagent in Claude Code (a text file with instructions instead of clicks), or in Claude Cowork (for tasks in your folders, without a terminal). Which way fits depends on whether the AI only needs to answer or needs tools: reading and changing files, running commands, handing work to other agents. Way 1 is the preliminary stage and the right place to start: role, instructions, and knowledge, without a line of code, first usable result after about an hour in my experience. It becomes an agent on ways 2 and 3, where tools and self-directed steps come in. What the three ways cost, and where each one stops, is further down.
My entire AI workforce runs on way 2: each employee is a folder with instructions, knowledge, and skills, and can run as a subagent in Claude Code. The fundamentals are on the AI agents page; how subagents, skills, and agents differ is in Claude Subagents, Skills, Agents.
Build an AI agent with Claude: three ways compared
The prices, with date and source, are in Claude Pricing and Plans Explained. I describe way 1 in the most detail because its five steps apply to the other two as well.
Way 1: the preliminary stage as a Claude project in five steps
A Claude account is enough. According to Anthropic's pricing page, projects are available on the free plan too, with tight message limits once you work daily. More important than the plan is the task: it repeats, you can judge its result in minutes, and it costs you time. I'll use "preparing proposals" as the example.
Step 1: Create a project. Create a new project in Claude and name it after the job, for example "Proposal Agent." According to Anthropic's help page on projects, a project has two properties that turn any chat into a fixed role: its own instructions, which apply to every conversation inside it, and a knowledge base every conversation can access. One project per job; a jack-of-all-trades is nearly impossible to correct.
Step 2: Write the instructions. In the project instructions, write what you'd tell a new hire on their first day:
- The job: "You prepare proposals. You get my conversation notes and deliver a finished proposal draft."
- The process: "First clarify services and scope, then prices according to my price list, then the text based on my examples."
- The boundaries: "You never invent prices or services. If you're missing a piece of information, you ask instead of guessing."
- The format: "Deliver a finished document, structured like my example proposals."
The point under "boundaries" matters most: an AI that asks instead of guessing saves you the rework.
Step 3: Upload knowledge. Now fill the knowledge base with what the role needs for the job: your price list, two or three real proposals as style templates, a short description of your business, and phrases you would never use. A trick from my practice: in the first conversation, write "Ask me all the questions you need for your job," answer them, and put the answers into the knowledge base as a document.
Step 4: Test on a real case. Use the proposal that has to go out this week anyway. Don't rewrite mistakes yourself, name them: "The price on line item 2 is wrong," "too formal, look at example 3." Every correction you silently make on your own is one Claude will get wrong again on the next case.
Step 5: Make corrections permanent. Every correction that should apply more than once goes back into the instructions or into a document in the knowledge base. "Too formal" becomes one line: "You write casually, no bureaucratic language." After a few rounds, every new case is routine: notes in, draft out, you approve.
An honest word on the limit of this way: I've used Claude Projects myself and never got comfortable with them. Everything sits inside the project at the provider, and there is no way out again, not to Codex and not to an open model. To me that feels like less control. It's still a good way to start, and for pure text work it stays the fastest.
Way 2: an agent in Claude Code as a subagent
In Claude Code, an agent is a Markdown file. According to Anthropic's subagents documentation, it lives under .claude/agents/ in the project folder or under ~/.claude/agents/ for all your projects. The file's header (frontmatter) needs two required fields, name and description; below it, in plain prose, you write what the agent should do. tools optionally limits what it may use. An example for the same job:
---
name: proposal-check
description: Checks finished proposal drafts against the price list and style templates. Use as soon as a proposal draft appears in the proposals/ folder.
tools: Read, Grep, Glob
---
You check proposal drafts. Read the price list in knowledge/prices.md
and the style templates in knowledge/examples/. Report every deviation
with its location. You change nothing yourself, you only report.
You call the subagent from your running session, either because Claude decides from the description that it fits, or directly, with @agent-proposal-check or in plain language ("Use the proposal-check subagent on today's draft"). It works in its own, empty context: without your conversation so far, but with your project's CLAUDE.md and its task text. Only its summary comes back. That's why my workforce is built this way: the reviewer didn't watch the draft being written, it sees it fresh.
The difference from way 1 is the tools: a subagent can read and write files, run commands, and reach external systems through MCP servers, such as a mailbox; how to set that up is in Set Up a Claude MCP Server. With tools: Read, Grep, Glob as in the example, it may only read. That's my rule for every reviewing role: whoever reviews changes nothing.
Way 3: Cowork for tasks with files
According to Anthropic's getting-started page for Cowork, Claude Cowork uses the same agent architecture as Claude Code, just without a terminal. You share folders, describe the task in plain language, and Claude works in exactly those folders: formatting documents, building spreadsheets, merging research. Every step is shown, and the result lands in your folder for review. Since September 16, 2026, Anthropic has been merging Cowork and chat into one Claude, rolling out gradually, starting with Pro and Max. After the switch you no longer pick Cowork separately; you describe the task in a normal conversation, and until then the message box still shows the "Cowork" option. You share folders on your computer in Claude Desktop, and Claude only reaches them while the app is open (support.claude.com, as of October 8, 2026).
I built Skills for Cowork and uploaded them as a ZIP; that worked well. My own day still runs almost entirely in Claude Code, and has for about three months. Cowork fits when your agent needs to touch files but you don't want a terminal; for several agents with fixed rules, way 2 is the more direct route. What Cowork can do, what it costs, and where the line to Claude Code runs is in Claude Cowork: What It Does, Who It's For.
Claude agent teams: letting several agents work together
With more than one agent, the question is how they hand work to each other. For me, that runs through a mailbox protocol: the sender puts a handover file in its outbox and a copy in the recipient's inbox, with a link to the context instead of a copy of the content. Every employee checks its mailbox at the start of a session, and everything that leaves the house is reviewed by a separate role.
One sequence from daily operations, September 7, 2026: Sebastian, my SEO employee, writes Ralf, who builds and writes websites, the handover for a blog release. I approve at 18:24. Ralf works as a subagent because delivery to his running session failed at the approval step. Sebastian then reviews the result live. On September 16 it ran the other way round: Sebastian asks Conny, my content editor-in-chief, through the mailbox for experiences for five posts, and Conny answers as a subagent with a letter in his inbox. What Sebastian makes of input like that is on Sebastian, who measures my website with real tools.
Two rules hold this together. First: an agent commissions another agent only on my instruction; Conny's letter to Eddi, my video employee, then carries my exact wording, context links, mandatory checks, and an approval gate. What becomes of such a letter on Eddi's side is on Eddi, who cuts my reels from raw footage. Second: approvals come from me alone; the employee documents them with a timestamp. What such an interplay looks like for bookkeeping and inbox is in AI Agent for Bookkeeping and Inbox. Which patterns hold up from the second agent on, and who may see and change what, is laid out step by step in Orchestrating Multiple AI Agents.
What Claude agents cost
For way 1, a Claude account is enough; Free costs 0 dollars, Pro 20 dollars a month (as of September 2026, source: Anthropic's pricing page). Ways 2 and 3 need a paid plan: Pro 20, Max 100 or 200 dollars a month, Team from 20 dollars per seat a month. I didn't start on Max; today I run Claude Team with two seats plus Max, Team for business data because of the data processing agreement. All tiers with sources are in Claude Pricing and Plans Explained, and whether a plan or the API is cheaper for Claude Code is in Claude Code Pricing: Plan or API?.
From agent to employee
After way 2 or 3 you have a working AI agent: a job, context, tools, and a repeatable result; after way 1 you have a fixed role with your knowledge behind it. What both are missing is the management layer: a name, a proper personnel file instead of scattered documents, a clean onboarding, a memory for corrections, and feedback as a fixed part of the work. What exactly that means is in What Is an AI Employee?.
That's a shift in perspective, not a technical step: you manage instead of operating, and you remain the one who approves. The complete process from choosing the task to the permanent hire is in Hiring an AI Employee: The Process.
Frequently asked questions
Do I need Claude Code, or is Claude.ai enough?
For a role that delivers text and answers from your knowledge, a project in Claude.ai is enough. You need Claude Code as soon as the AI should work in files itself, run commands, or commission other agents; from there on I call it an agent.
Does this work on the free plan?
Way 1 yes, with tight message limits. Ways 2 and 3 no; Claude Code and Cowork need a paid plan. What the free plan can do and when Pro pays off is in Build an AI Agent for Free: What Works?.
Agent or project, what's the difference?
A project is the container: instructions plus a knowledge base that apply in every conversation. With a clear job and corrections on real cases it becomes a fixed role, but technically still a chat with context. It becomes an agent with tools and self-directed steps, so in the subagent or in Cowork.
What is the agent mode?
I haven't found a switch with that name in Anthropic's docs. What people usually mean is that Claude plans several steps and uses tools instead of just answering. That's what Claude Code or Cowork give you; a project in Claude.ai remains a chat with context.
How do I build a specialist agent, for example for bookkeeping?
Like the proposal agent above, with different knowledge: your chart of accounts, your receipt rules, examples of correctly booked cases. The key instruction stays "ask instead of guessing": on September 6, 2026, my bookkeeping employee Peter, working through 63 receipts, raised two questions before producing output, one about two currencies, instead of inventing exchange rates. That he only does what his written process describes is his ground rule: Peter, my employee for bookkeeping.
Can the agent read emails?
Only with a connection: in Claude.ai through one Anthropic offers, in Claude Code through an MCP server for your mailbox. Reading and drafting is the easy part; sending stays behind an approval for me, because everything that leaves the house is reviewed first.
Where to go from here
Take a real task and way 1. When the agent delivers cleanly after a few rounds and you notice it would need to touch files, that's the moment for way 2 or 3. Whether Claude or ChatGPT fits your agent better is in Build an AI Agent: Claude vs ChatGPT, and ten agents from daily operations, each with a result, are in AI Agents: 10 Real-World Examples.
And if you're wondering why I build agents in Claude rather than in n8n or Make: the reasoning, with a comparison table, is in Building AI Agents Without Code: My Way. The short version: workflow platforms solve assembly-line tasks, but the expensive hours of a solo business sit in tasks that need context and judgment. And if you are missing the order to get there, you will find it in Learn Claude: The Path in Five Stages.
