How to Build AI Agents With Claude in 2026 Using Four Practical Workflows
Video: How To Build $250,000 AI Agents With Claude (Step-By-Step)
· based on the channel Charlie Chang
Key takeaways
- Claude Code agents can be built by non-developers; the workflow is described step by step, starting from a plain-language brief.
- The video walks through four agents: a content strategist, an executive assistant for email, a lead generation agent and a customer support agent.
- Zapier MCP connects Claude to 8,000+ apps, including Gmail and Google Drive, so agents can read and write in real accounts.
- Running Claude Code on a Bluehost VPS keeps agents working when the laptop is off, which is required for 24/7 support and lead scraping.
- A viewer reported that the first content agent did not ask how much content to produce per video, a sign the brief or skill file needs explicit output rules.
Building AI agents with Claude means giving the model a written brief, a set of tools it can call, and a place to run continuously; Claude Code handles the first two, and a small VPS handles the third. In practice you describe the job in plain language, connect the accounts the agent needs through MCP, and let it return finished work instead of chat replies. The hosting link used in the walkthrough, a Bluehost VPS for running Claude Code automations 24/7, covers the always-on part.
What separates an AI agent with Claude from a chatbot
A chatbot answers; an agent executes a defined job and hands back a result. The difference is structural, not a matter of a better prompt: an agent needs a role, a repeatable input, tool access, and an output format it can be judged against.
Claude Code is the practical entry point because it runs in a terminal-like environment where the model can read files, write files and call external services. Four ingredients make that useful:
- A brief that states the role, the input and the exact deliverable.
- Tool connections so the agent can reach Gmail, Google Drive, a CRM or a spreadsheet.
- A trigger, either manual or scheduled, that starts the run.
- A destination where the output lands, such as a draft folder or a review queue.
Skip any one of these and you get a demo rather than a working agent. The most common failure is a vague deliverable: if the brief does not say how many outputs to produce, the agent will guess or ask nothing at all.
The four agents built in the walkthrough
Charlie Chang builds four agents from scratch, each aimed at a task a small team normally does by hand. They escalate in complexity, which makes them a sensible learning order.
- Content strategist. Takes a single video or long-form piece and turns it into a week of content: repurposed posts, captions and a posting plan. This is the simplest agent because the input is one file and the output is text.
- Executive assistant. Sorts incoming email, labels what matters and drafts replies for review. It needs a live mailbox connection, which is where MCP enters.
- Lead generation agent. Finds businesses that match a profile and writes personalized outreach for each one. This agent combines research, filtering and writing, so it benefits most from running on a server.
- Customer support agent. Answers recurring questions around the clock. It only works if it runs somewhere that never sleeps, because a laptop that closes is a support desk that closes.
The progression is deliberate: text in, text out first; then one connected account; then research plus outreach; then a public-facing agent with uptime requirements.
Connecting Claude to Gmail, Drive and 8,000+ apps with MCP
MCP, the Model Context Protocol, is the standard that lets Claude call outside tools without custom code for each one. Instead of writing an integration per service, you connect an MCP server and the agent gains a defined set of actions.
Zapier MCP is the shortcut used in the video, exposing 8,000+ apps through one connection. The setup pattern is the same regardless of the tool:
- Create the connection in the MCP provider and authorize the specific account.
- Grant the narrowest scopes that still let the agent do the job, for example read and draft rather than send and delete.
- Add the connection to the Claude Code environment so the agent can see the available actions.
- Test with a single record before pointing the agent at a full inbox or a large lead list.
For the executive assistant, this means the agent can read new mail, apply labels and save draft replies. For the lead agent, it means writing rows into a sheet or a CRM. Keep sending rights off until you have reviewed a batch of drafts; an agent with send permission and a bad brief is an incident, not an automation.
Running Claude Code on a VPS for 24/7 agents
The customer support agent and the lead generation agent both need uptime, and a local machine cannot provide it. A virtual private server gives Claude Code a permanent home with its own operating system, storage and scheduled tasks.
The video's setup uses a Bluehost VPS, with a separate document covering the full installation steps. The general sequence is:
- Provision the VPS and connect over SSH.
- Install the runtime and Claude Code, then authenticate the account.
- Move the agent's brief and skill files onto the server so they persist between runs.
- Schedule the runs, for example an inbox pass every morning and a support check at short intervals.
- Log every run so you can see what the agent did and where it failed.
Two practical notes. First, cost scales with activity, so a support agent checking every minute burns far more than one checking every fifteen. Second, keep secrets in environment variables on the server rather than inside prompt files, because briefs get copied and shared.
Common problems when your first agent ignores instructions
A frequent report from people following the tutorial is that the first content agent never asks how much content is wanted from each video. That is not a bug in Claude; it is a gap in the brief.
Agents do not reliably ask clarifying questions, so the instructions must state the output count, the length of each piece, the platform and the tone. A useful fix is to add an explicit output contract: "produce one LinkedIn post of 150 words, three short posts, and one email subject line per video." Then test with a single input and check the count before scaling.
Other recurring issues and their causes:
- The agent answers generically. The brief lacks context files, so add a document describing the business, audience and offers.
- Tools do not appear. The MCP connection was authorized for a different account or the scopes are too narrow.
- Runs stop overnight. The process is running locally; move it to the VPS and confirm the scheduler fires.
- Output quality drifts. Add two or three examples of good output to the skill file and require the agent to match their structure.
Treat the first week as calibration. Review every output, tighten one instruction at a time, and only then remove the human check.
Links and resources from the walkthrough
All resources referenced in the breakdown, copied as published:
- Bluehost VPS for running Claude Code automations 24/7: https://bluehost.sjv.io/rENbY3
- Full VPS setup instructions: https://docs.google.com/document/d/1PSm8hP0wH6D1oOde_QgXqo-LHcyq5xBUamOmcj9Ii6g/edit?usp=sharing
- Bluehost All Access, one plan for Claude, ChatGPT, Gemini and Grok: https://charliechang.com/bluehostallaccess
- Free AI community with templates and workflows: https://www.skool.com/ai-os/
- Zapier MCP, connecting Claude to 8,000+ apps: https://bit.ly/4vPaX9x
Start with one agent, then compound the savings
The value of AI agents with Claude is not the first automation but the pattern it teaches. Once a content agent turns one video into a week of posts, the same structure applies to invoices, research summaries, onboarding emails and support macros. Charlie Chang's breakdown is a useful map because it moves from a single-file agent to connected accounts and finally to always-on hosting, which is the order most people should follow. Begin with the content strategist, add one MCP connection, and only move to the VPS when a task genuinely needs to run while you sleep. When you are ready for that step, get a Bluehost VPS to keep Claude Code agents running 24/7 and schedule your first overnight job.
Questions & answers
Do I need to know how to code to build AI agents with Claude?
No. The walkthrough builds all four agents from written instructions and tool connections, not application code. You do need basic comfort with a terminal and with copying files to a server, because Claude Code runs in that environment. The VPS setup document covers the installation steps in order.
Why did my content agent not ask how much content I want per video?
Agents rarely ask clarifying questions on their own, so the output rules must be in the brief. State the number of pieces, the length of each, the platform and the tone explicitly, then test with one input and check the count. If the count is wrong, edit the instruction rather than re-prompting each time.
What is the difference between Zapier MCP and running Claude Code on a VPS?
MCP is the connection layer that lets Claude read and write in outside apps such as Gmail, Google Drive or a CRM. A VPS is the machine the agent runs on, so it keeps working when your laptop is closed. The lead generation and customer support agents need both: tools to act and a host that never sleeps.
How much does it cost to keep an agent running around the clock?
The main costs are the VPS subscription and the model usage, which scales with how often the agent runs and how much it reads. A support agent checking every minute costs far more than one checking every fifteen minutes. Start with a modest schedule, review the logs, and increase frequency only where response time actually matters.
Source: How To Build $250,000 AI Agents With Claude (Step-By-Step) · Markdown version
