1
Starter
Your First Agent · 1–2 hrs
Build a working AI automation using zero code — just Claude and Zapier. Your agent will respond to triggers, answer questions, and take actions automatically. No terminal, no installation, no experience needed.
Claude Free
Zapier
100% Free to Start
No Code
No Installation
💡 What you can build with no-code agents
Email auto-responder
Form → AI → Slack reply
Daily briefing emailer
Lead qualifier bot
Meeting notes summarizer
Social post generator
Support ticket tagger
Document classifier
Newsletter curator
Customer feedback analyzer
Invoice data extractor
RSS → summary digest
🔓 Skills you'll unlock
💡 Stuck on a step? Paste the error or describe what's not working into Claude.ai or your Claude Code terminal — it will diagnose and fix it. Getting unstuck this way is a core part of the process.
Understanding API Costs — Before You Build Anything
▾
This is the most important thing to understand before building agents. Unlike web apps that run on free Vercel hosting, agents call APIs every time they run — and those calls cost money. Understanding costs upfront prevents nasty surprises.
1
How Claude API pricing works
You pay per token — roughly per word. Claude Sonnet 4 costs approximately $0.003 per 1,000 input tokens and $0.015 per 1,000 output tokens. A typical agent call costs about $0.01–$0.05. That's nearly free for testing, but adds up if your agent runs thousands of times. Check current pricing at anthropic.com/pricing.
2
Estimate your agent's monthly cost before deploying
The formula:
(runs per day) × (cost per run) × 30. An agent that runs 24 times per day (hourly) at $0.02 per run = $14.40/month. One that runs every 5 minutes at $0.05 = $432/month. Always calculate this before setting a schedule.3
Set a spending limit before you start
In the Anthropic Console, go to Billing → Usage Limits and set a monthly hard limit — e.g. $20. This caps your exposure if an agent runs in an unexpected loop. Set it before writing a single line of agent code.
4
Keep prompts short and efficient
Every word in your system prompt costs money, multiplied by every run. Cut anything that isn't necessary. A 2,000-token system prompt running hourly costs 10× more than a 200-token one. Ask Claude: "Rewrite this system prompt to be as concise as possible while keeping all the important instructions."
Prompt Engineering for Agents — Different from Chatting
▾
Writing prompts for an autonomous agent is very different from chatting with Claude. In a chat, you can clarify and correct. In an agent, the prompt runs unattended — every ambiguity becomes a bug.
1
Give the agent a clear role and scope
Start every agent system prompt with: what the agent is, what it should do, and — crucially — what it should NOT do. Constraints are as important as instructions.
You are an email response assistant for [Business Name].
Your job: read incoming customer support emails and draft a helpful reply.
ONLY respond to questions about our products and services.
DO NOT make promises about refunds, timelines, or pricing without checking.
If you're unsure how to respond, say "I'll need to check on this" instead of guessing.
2
Specify the output format exactly
Agents pass their output to the next step — a webhook, an email, a database. If the format varies, downstream steps break. Tell the agent exactly what to return: "Respond with only a JSON object with keys: reply_text (string), confidence (1-10), needs_human_review (boolean). No other text."
3
Test edge cases before going live
Before turning on an automated agent, manually test at least 5 edge cases: empty input, very long input, input in another language, rude or nonsensical input, and the exact type of input you expect most. Ask Claude: "What are 5 edge cases that might cause this agent to behave unexpectedly? Let's test each one."
What You're Building — and What an Agent Actually Is
▾
An agent is an AI that doesn't just answer questions — it takes action. Instead of you typing a prompt and reading a response, an agent watches for a trigger (like a new email or a form submission), processes it using AI, and does something with the result (like sending a reply, updating a spreadsheet, or posting to Slack) — all without you touching a keyboard.
What you'll build: A Zapier workflow that watches your Gmail inbox for emails with the subject "AI Help", feeds the email body to Claude, and sends an intelligent AI-written reply automatically — within seconds of receiving the email.
Step 1 — Set Up Your Accounts
▾
1
Get Claude API access
Go to console.anthropic.com and sign up. You pay only for what you use — testing typically costs cents. Add a small credit (e.g. $5) to get started.
2
Create a free Zapier account
Go to zapier.com and sign up. The free plan allows 100 tasks per month — more than enough to build and test your first agent.
3
Get your Anthropic API key
In the Anthropic Console, go to API Keys → Create Key. Copy it somewhere safe — you'll paste it into Zapier. Never share this key publicly or commit it to GitHub.
Step 2 — Build Your First Zap (Agent Workflow)
▾
1
Create a new Zap in Zapier
In Zapier, click Create Zap. A Zap is a workflow — it has a trigger (what starts it) and one or more actions (what happens next).
2
Set the trigger — New email in Gmail
Search for Gmail as your trigger app. Choose the event "New Email". Connect your Gmail account. Under filters, set Subject Contains to
AI Help. This means the Zap only fires when you receive an email with that subject line.3
Add Claude as the action — generate a reply
Click the + to add an action. Search for Anthropic (Claude). Choose "Send Message". Connect using your API key. In the prompt field, enter:
You are a helpful assistant. Someone emailed asking for help. Read their email and write a friendly, helpful reply in 2-3 sentences.
Email subject: [insert Gmail subject field here]
Email body: [insert Gmail body field here]
Write only the reply text — no subject line, no greeting like "Dear", just the reply body.
Use Zapier's variable picker to insert the actual Gmail data from the trigger step into the prompt.
4
Add a second action — send the reply
Add another Gmail action. Choose "Send Email". Set To to the sender's email from the trigger. Set Subject to
Re: [original subject]. Set Body to Claude's output from the previous step.5
Test and turn on your agent
Click Test in Zapier to run the workflow with a real email. Review the output — does Claude's reply make sense? Adjust your prompt if needed. When satisfied, toggle the Zap to On. Your agent is now live.
Test it: Send yourself an email with the subject "AI Help" and a question in the body. Within 30 seconds you should receive an AI-written reply.
Step 3 — Customize and Extend
▾
Now that your basic agent works, try these improvements:
- Change the trigger — instead of Gmail, try a Typeform submission, a Slack message, or a new row in Google Sheets
- Change the action — instead of replying by email, post Claude's response to a Slack channel or save it to a Google Doc
- Improve the prompt — give Claude a persona: "You are a customer support agent for [your business]. Always be friendly, concise, and offer one next step."
- Add a filter step — only trigger the agent if the email contains certain keywords, or only during business hours
- Add logging — add a Google Sheets action that saves every email + Claude's reply so you can review them
2
Beginner
File Intelligence Agent · 1–3 hrs
Build a local AI assistant using Claude Code that reads documents in a folder and answers questions about them in plain English. Drop any file in — PDFs, contracts, reports, spreadsheets — and ask anything.
Claude Code
Python
File Processing
Local Agent
RAG Basics
🔓 Skills you'll unlock
💡 Stuck on a step? Paste the error or describe what's not working into Claude.ai or your Claude Code terminal — it will diagnose and fix it. Getting unstuck this way is a core part of the process.
What You're Building
▾
How it works: You run a Python script from your terminal. It reads all documents in a
docs/ folder. You type a question. It sends the documents to Claude along with your question, and gives you an answer based only on what's in those files.- Ask questions about any PDF, .txt, or Word document
- Compare multiple documents at once
- Extract specific data — dates, names, numbers, clauses
- Summarize long documents in seconds
- Find contradictions or inconsistencies across files
Step 1 — Set Up and Build with Claude Code
▾
1
Create a project folder and start Claude Code
mkdir file-agent
cd file-agent
mkdir docs
claude
2
Describe the agent to Claude Code
"Build me a file intelligence agent in Python. Here's what it should do:
1. Watch a folder called 'docs/' in the current directory
2. When I run the script, it reads all .txt, .pdf, and .docx files in that folder
3. It presents a simple command-line interface where I can type questions
4. It sends the contents of all the files plus my question to the Anthropic API
5. It prints Claude's answer in the terminal
6. It keeps running so I can ask multiple questions — type 'quit' to exit
Use the Anthropic Python SDK. Store the API key in a .env file. Add clear comments explaining how each part works. Keep the code as simple as possible."
3
Install dependencies and create your .env file
"Install all required Python packages and create a .env file template I can fill in with my Anthropic API key."
Open the .env file and replace the placeholder with your actual API key from console.anthropic.com.
4
Drop files in docs/ and test
Copy any PDF, Word doc, or text file into the
docs/ folder. Then run the agent:
python agent.py
Type a question like "What are the main points in these documents?" and press Enter.
Step 2 — Extend Your Agent
▾
Once the basic agent works, ask Claude Code to add these improvements one at a time:
- Conversation history — remember previous questions in the same session so follow-up questions make sense
- File summaries on startup — when the agent loads, automatically summarize what each document contains
- Source citing — have Claude name which document each piece of information came from
- Save Q&A to a log file — save every question and answer to a timestamped text file for reference
Example prompt: "Add conversation history so Claude remembers the last 5 questions and answers in each session. This way I can ask follow-up questions like 'tell me more about that' and it will understand."
3
Intermediate
Web Research Agent · 2–4 hrs
Build an agent that searches the web on a topic you choose, reads the results, synthesizes the findings, and emails you a clean formatted digest — automatically, on a schedule.
Claude Code
Python
Serper API
Web Scraping
Email Automation
Scheduling
🔓 Skills you'll unlock
💡 Stuck on a step? Paste the error or describe what's not working into Claude.ai or your Claude Code terminal — it will diagnose and fix it. Getting unstuck this way is a core part of the process.
Error Handling — Agents Fail Silently Without This
▾
An agent running unattended can fail for hours without anyone noticing. Without error handling, you'll never know. With it, failures surface immediately.
1
Wrap every external call in try/except
Ask Claude Code: "Add try/except error handling around every external API call. If a call fails, log the error with a timestamp and continue rather than crashing the whole agent."
2
Validate Claude's output before using it
When your agent expects Claude to return JSON, always validate it before passing it downstream. Ask Claude Code: "Before using Claude's response, check that it matches the expected format. If it doesn't, retry once. If it fails again, log the bad output and skip this run."
3
Add a dead agent alert
Ask Claude Code: "If the agent hasn't successfully completed a run in 3 hours, send me an alert email. This catches situations where the agent crashes silently." This one feature will save you countless headaches.
4
Handle empty or poor search results gracefully
Never send Claude an empty results set — it will hallucinate content that doesn't exist. Always check: "If the search returns fewer than 2 results, skip this run and log 'insufficient results' rather than generating a digest."
What You're Building
▾
The pipeline: You define a topic (e.g. "AI news" or "competitor mentions"). Every morning the agent searches the web for recent articles, reads the top results, passes them to Claude to synthesize into a 5-bullet digest, then emails it to your inbox. Zero manual effort.
- Uses Serper API (free tier: 2,500 searches/month) to find current articles
- Fetches and reads the actual page content — not just headlines
- Claude synthesizes across multiple sources into a coherent digest
- Formats the email with clear sections, links, and key takeaways
- Runs automatically every day via a scheduler
Step 1 — Get Your API Keys
▾
1
Get a free Serper API key (web search)
Go to serper.dev and sign up. The free plan gives you 2,500 searches per month — more than enough for daily research. Copy your API key.
2
Set up email sending
Use Gmail with an App Password. In your Google account, go to Security → 2-Step Verification → App Passwords. Create an app password for "Mail". This lets Python send email on your behalf without your real password.
Step 2 — Build the Research Agent
▾
1
Create a project and start Claude Code
mkdir research-agent
cd research-agent
claude
2
Build the agent pipeline
"Build a web research agent in Python with these steps:
Step 1 — Search: Use the Serper API to find the 5 most recent articles about a topic defined in a config variable at the top of the file
Step 2 — Fetch: For each result, fetch the actual page content using requests and BeautifulSoup, extracting just the main article text (no nav, footer, ads)
Step 3 — Synthesize: Send all article content to Claude with this task: summarize the most important findings into a 5-bullet digest with a one-line intro. Include the source URL for each bullet.
Step 4 — Email: Send the formatted digest to an email address defined in config, using Gmail SMTP with an app password
Store all API keys in a .env file. Add a config section at the top with: topic, my email, and how many articles to fetch."
3
Test it manually first
python research_agent.py
Check your inbox within 60 seconds. If the email arrives and the digest looks good, schedule it.
4
Schedule it to run every morning
"Add a scheduler so this script runs automatically every day at 7am. Show me how to set this up on my operating system — explain both Mac (cron) and Windows (Task Scheduler) options."
4
Advanced
Multi-Agent Pipeline · 3–5 hrs
Build a system of three specialized AI agents that hand off work to each other: a Researcher finds information, a Writer turns it into polished content, and a Reviewer critiques and improves it — all chained together automatically.
Claude Code
Python
Multi-Agent
Agent Orchestration
Prompt Chaining
🔓 Skills you'll unlock
💡 Stuck on a step? Paste the error or describe what's not working into Claude.ai or your Claude Code terminal — it will diagnose and fix it. Getting unstuck this way is a core part of the process.
What You're Building
▾
The pipeline: You give the system a topic. The Researcher searches and gathers facts. It passes those to the Writer, which drafts a 500-word article. The Writer's output goes to the Reviewer, which scores the piece and rewrites any weak sections. Final output is saved as a formatted markdown file.
- Researcher agent — searches the web, extracts key facts, cites sources
- Writer agent — receives the research, writes a structured article in a specified style
- Reviewer agent — reads the draft, scores it on clarity/accuracy/flow, rewrites anything below threshold
- Each agent has its own system prompt defining its role and constraints
- The pipeline logs what each agent did and why at every step
Step 1 — Build the Pipeline
▾
1
Create a new project and start Claude Code
mkdir multi-agent
cd multi-agent
claude
2
Build the three-agent system
"Build a multi-agent content pipeline in Python with three specialized agents, each using the Anthropic API:
RESEARCHER AGENT
- System prompt: You are a research specialist. Given a topic, search the web using Serper API and extract the 5 most important facts with source URLs. Return structured JSON: {facts: [], sources: []}
WRITER AGENT
- System prompt: You are an expert writer. Given research facts, write a clear, engaging 500-word article. Structure it with a headline, 3 sections, and a conclusion. Return as markdown.
REVIEWER AGENT
- System prompt: You are an editorial reviewer. Score the draft 1-10 on clarity, accuracy, and flow. If any score is below 7, rewrite that section. Return: {scores: {}, final_article: ''}
ORCHESTRATOR
- Takes a topic as input → runs Researcher → passes to Writer → passes to Reviewer → saves final article to output/[topic].md
Add detailed logging so I can see what each agent is doing at every step."
3
Run the pipeline on a test topic
python pipeline.py "the future of solar energy"
Watch the logs as each agent runs. Check
output/ for the finished article. The whole pipeline should take 30–60 seconds.
Step 2 — Add Retry Logic and Quality Gates
▾
"Add quality gates to the pipeline:
1. If the Reviewer scores any section below 6, automatically send it back to the Writer with specific feedback and retry up to 2 times
2. If the Researcher returns fewer than 3 facts, retry the search with a broader query
3. Log all retries with the reason so I can see when and why the pipeline self-corrected"
5
Pro
Agent with Memory · 4–6 hrs
Build an AI assistant that remembers. It learns your preferences, recalls past conversations, and gets more personalized every time you use it — powered by a real database and vector search that persists between sessions.
Claude Code
Python
Supabase
Vector Embeddings
Long-Term Memory
MCP Servers
🔓 Skills you'll unlock
💡 Stuck on a step? Paste the error or describe what's not working into Claude.ai or your Claude Code terminal — it will diagnose and fix it. Getting unstuck this way is a core part of the process.
MCP Servers — Give Your Agent Native Tool Access
▾
The Model Context Protocol (MCP) is an open standard that lets Claude connect directly to external tools and data sources — without you writing custom API integration code. Instead of building a custom Slack connector, you install a Slack MCP server and Claude can read and write to Slack natively. Claude Code has built-in MCP support, making this very straightforward to set up.
1
What MCP unlocks for agents
With MCP servers, your agent can connect to Google Drive, GitHub, Slack, Notion, databases, file systems, and dozens of other services using pre-built connectors. This dramatically reduces the work needed to build powerful agents that interact with your real tools.
2
Finding and installing MCP servers via Claude Code
Claude Code can discover, install, and configure MCP servers for you. To add one:
"I want to connect my agent to [Notion / GitHub / Slack / Google Drive]. Find the appropriate MCP server, install it in my Claude Code environment, and configure it. Walk me through getting the required API credentials and setting up the connection."
3
Add MCP-powered tools to your memory agent
For this project, use an MCP server to give your memory agent access to your calendar, notes, or email — so it learns from your actual digital life, not just conversations:
"Connect my memory agent to my Google Calendar via MCP. When I ask 'what do I have coming up?', the agent should check my actual calendar and incorporate it into its response."
What You're Building
▾
The problem with most AI: Every conversation starts from zero. It doesn't know your name, your preferences, or what you talked about last Tuesday. This project fixes that by giving your agent a persistent memory stored in a real database.
- Every conversation is stored as a vector embedding in Supabase
- When you ask something new, the agent searches memory for relevant past context
- It builds a preference profile — your name, communication style, topics you care about
- Each session gets smarter — it references past conversations naturally
- Memory is searchable: "What did I ask you about last week?"
Step 1 — Set Up Supabase with Vector Support
▾
1
Create a Supabase project and enable pgvector
Go to supabase.com and create a new project. In the SQL editor, run:
create extension if not exists vector;
2
Ask Claude Code to set up the database schema
"I'm building an agent with persistent memory using Supabase and pgvector. Create the SQL to set up:
1. A 'memories' table with: id, content (text), embedding (vector 1536), created_at, session_id, importance_score
2. A 'user_profile' table with: id, key (text), value (text), updated_at — for storing preferences
3. A function to search memories by semantic similarity using cosine distance
Give me the SQL to run in Supabase's SQL editor."
Step 2 — Build the Memory Agent
▾
1
Build the agent with memory capabilities
"Build a memory-powered AI assistant in Python. It should:
BEFORE each response:
- Convert my message to an embedding using the Anthropic embeddings API
- Search Supabase for the 3 most semantically similar past memories
- Retrieve relevant user preferences from the profile table
- Include all of this as context in the system prompt
DURING the conversation:
- Have a natural conversation using Claude as the AI
- Detect and extract any preferences I express (e.g. 'I prefer concise answers')
AFTER each exchange:
- Save the conversation turn as a new memory with an embedding
- Update user_profile if new preferences were detected
- Score each memory for importance (1-10) based on content type
The agent should feel personal — naturally reference past context when relevant."
2
Test memory persistence across sessions
Run the agent, have a conversation, close it, restart it, and ask: "What do you know about me?" or "What did we talk about last time?" The agent should recall your past conversations accurately.
6
Ninja
Autonomous Business Agent · 5–9 hrs
Build a fully autonomous agent that monitors a business trigger, decides what action to take, executes it, logs the result, and alerts you — running 24/7 in the cloud without any human involvement.
Claude Code
Python
Railway
Tool Use
Autonomous Execution
Cloud Deployment
🔓 Skills you'll unlock
💡 Stuck on a step? Paste the error or describe what's not working into Claude.ai or your Claude Code terminal — it will diagnose and fix it. Getting unstuck this way is a core part of the process.
Evaluating Your Agent — How to Know If It's Actually Working
▾
A deployed agent that's quietly making bad decisions is worse than a broken one. Production agents need evaluation — a way to measure whether they're doing their job well and catching when quality degrades.
1
Define what "good" looks like before you deploy
For each action your agent takes, decide upfront what success looks like. Write these criteria into your code as assertions — if the agent's output doesn't meet them, it logs a quality warning.
2
Build a review log you actually read
Ask Claude Code: "Add a weekly summary email showing: total runs, success rate, number of actions taken, quality warnings triggered, and 3 example outputs. Send it every Monday morning." An agent you never inspect is one you can't improve.
3
Detect when the agent's quality has changed
Ask Claude Code: "Add a quality score to each run — ask Claude to self-evaluate its output on a 1-10 scale. Log this score over time and alert me if the 7-day average drops below 7."
Trust Boundaries — What Your Agent Should Never Be Allowed to Do
▾
An autonomous agent that can take action in the real world needs hard limits — not soft guidelines, but code-enforced guardrails. Trust boundaries are the difference between a useful tool and an expensive mistake.
1
The principle of least privilege
Give your agent access to only what it needs. If it only needs to read emails, don't give it permission to send them. Scope credentials tightly and document what each one is for.
2
Hard-code irreversible action limits
For each irreversible action (sending emails, posting publicly, spending money), add a code-level rate limit the agent cannot override. Ask Claude Code: "Add hard rate limits: maximum 10 emails per hour, 50 Slack messages per day. If either limit is hit, stop all actions and alert me — no bypassing this."
3
Build a human-in-the-loop escape hatch
Ask Claude Code: "For any action with an importance score above 8, send me an approval request via email with a one-click approve/reject link. Only proceed if I approve within 24 hours."
What You're Building
▾
The agent loop: Every hour the agent checks a data source (e.g. brand mentions on Reddit, competitor job postings, a stock price). It uses Claude with tool calling to decide: is this worth acting on? If yes, it takes a defined action (email, Slack, log to spreadsheet). It runs this loop continuously — deployed in the cloud, costing ~$5/month.
- Monitors a trigger source every hour (configurable)
- Uses Claude's tool-use API — the agent can search, read URLs, send messages
- Claude decides whether each event is worth acting on based on criteria you define
- All actions are logged with timestamp, reasoning, and outcome
- Alert system notifies you when a high-priority event is detected
- Deployed to Railway — runs 24/7 in the cloud, survives computer restarts
Step 1 — Build the Autonomous Agent
▾
1
Create the project and build the agent
mkdir autonomous-agent
cd autonomous-agent
claude
"Build an autonomous business monitoring agent in Python. It should:
TOOLS (give Claude these capabilities via tool use / function calling):
- search_web(query): searches Google via Serper API, returns top 5 results
- fetch_url(url): fetches and extracts text content from a URL
- send_slack_message(message): posts to a Slack webhook URL
- send_email(subject, body): sends via Gmail SMTP
- log_event(event, action_taken, reasoning): appends to a local SQLite database
AGENT LOOP (runs every 60 minutes via a scheduler):
1. Fetch recent mentions of a monitored keyword from the web
2. For each result, ask Claude (with tools available): 'Is this worth acting on? If yes, what action should I take?'
3. Claude decides and calls the appropriate tool
4. Log every decision with full reasoning
SAFETY GUARDRAILS:
- Never take the same action twice on the same URL/content
- Maximum 10 actions per hour (rate limit)
- All actions require logging — nothing happens silently
MONITORING:
- A simple /status endpoint showing the last 20 events in the log
- Alert email if no events have been processed in 3 hours (dead agent detection)"
2
Test locally with a 5-minute interval
"Change the interval to 5 minutes temporarily so I can test the full loop quickly. Show me the log output as it runs so I can see each decision the agent makes."
Step 2 — Deploy to the Cloud
▾
1
Prepare for deployment
"Prepare this project for Railway deployment. Create:
1. A requirements.txt with all dependencies
2. A Procfile that runs the agent
3. A .gitignore that excludes .env and the SQLite database
4. Instructions for setting environment variables in Railway's dashboard"
2
Deploy to Railway
Go to railway.app, create a free account, connect your GitHub repo. Add all environment variables in the Railway dashboard. Deploy. Your agent starts running immediately in the cloud — 24/7, even when your computer is off.
You've done it. A fully autonomous AI agent running in the cloud, monitoring the world on your behalf, making decisions, taking actions, and alerting you — built entirely by describing what you wanted in plain English with Claude Code.