How to Build AI Agents with Plain English in 2026

The most valuable skill for a builder in 2026 isn’t coding; it’s directing AI agents that code for you. You don’t need to write Python or JavaScript. You need to write clear instructions in English. This post gives you the exact template I use to do it.

This matters now because the tools just caught up to the idea. With the launch of the Cursor SDK, anyone can describe what they want an application to do, and an autonomous agent can build, test, and run the code. I used this method to build a full email automation workflow in under 10 minutes. If you can talk it, you can build it.

---

What is an AI Agent?

An AI agent is a software program that can perceive its environment, make independent decisions, and take autonomous actions to achieve a specific goal. Unlike a simple script that follows a fixed path, an agent can reason, plan, and adapt its approach based on new information without direct human intervention.

---

The Agent Spec I Use — Copy It Right Now

This is a plain-language specification file that I feed directly to Claude Code to generate a working AI agent. This spec defines an agent that connects to an email inbox, reads unread messages, and generates a daily summary. You can copy this, change the details, and use it as the starting point for your own agent.

---
spec_name: "Daily Inbox Summarizer Agent"
version: "1.0"
author: "Mike Kwal"
goal: "An autonomous AI agent that connects to a specified IMAP email inbox, scans unread emails from the last 24 hours, and generates a daily summary with action items."
---

## 1. Purpose
This agent's job is to save time by eliminating the need to manually check and sort through a busy inbox every morning. It should provide a clear, actionable summary that can be read in 60 seconds.

## 2. Core Logic (Step-by-Step)
1.  **Connect:** Authenticate with the specified IMAP server using the provided credentials (host, user, password).
2.  **Fetch:** Search for all unread emails received in the last 24 hours.
3.  **Parse:** For each email, extract the sender, subject, and body text.
4.  **Analyze:** Use an LLM to determine the core topic and identify any explicit action items or questions in the email body.
5.  **Summarize:** Generate a single summary document for all processed emails.
6.  **Output:** Save the summary as a Markdown file to a local directory.
7.  **Mark as Read:** After successful processing, mark the emails as read on the server.

## 3. Inputs (Required)
- IMAP Host (e.g., `imap.gmail.com`)
- IMAP User (e.g., `[email protected]`)
- IMAP Password (use an app-specific password)

## 4. Output Format (Markdown)
```markdown
# Daily Email Summary - {YYYY-MM-DD}

## Action Items
- [ ] Reply to John Doe about the Q3 proposal.
- [ ] Send the invoice to Jane Smith.

## Key Information
- **Project Alpha:** The client approved the new mockups.
- **Team Update:** The weekly sync is moved to 3 PM.

## For Review (No Action Needed)
- Industry newsletter about AI trends.
```

## 5. Constraints & Edge Cases
- **Error Handling:** If the IMAP connection fails, log the error and stop. Do not mark emails as read.
- **No Unread Emails:** If no unread emails are found, create an empty summary file stating "No new messages today."
- **Security:** Do not log passwords or raw email content to the console.

To use this, I save it as `agent-spec.md` and give it to Claude Code with a simple prompt like: “Using the Cursor SDK and Node.js, build the agent defined in this spec file.”

+-----------------------+
|   Plain English Spec  |
|    (agent-spec.md)    |
+-----------+-----------+
            |
            v
+-----------+-----------+
|   Claude Code / Cursor|
+-----------+-----------+
            |
            v
+-----------+-----------+
|  Working AI Agent     | --> Monitors Inbox --> Daily Summary
|    (Node.js code)     |
+-----------------------+

---

Here’s Exactly How I Build an Agent

My process for building a new AI agent involves four distinct steps, moving from a high-level idea to a functional piece of software without writing the code myself. The goal is to stay in the director’s chair, defining the ‘what’ and letting the AI handle the ‘how.’ This keeps the development loop incredibly fast.

1. Define the Goal in One Sentence. I start by writing a single, clear sentence describing the agent’s job. For example: “Build an agent that monitors my inbox and summarizes action items every morning.” This clarity is the foundation for the entire build.
2. Write the Spec File. I take that one-sentence goal and flesh it out into a detailed spec, like the example asset above. I define the inputs, the step-by-step logic, the desired output format, and any rules or constraints. This spec becomes the single source of truth for the AI.
3. Generate the Code with an AI Tool. I drop the spec file into an AI development environment like Cursor or Claude Code. My prompt is direct: “Build the agent described in this spec file using Node.js.” The tool reads my instructions and generates the complete codebase.
4. Test and Refine the Instructions. I run the generated code. If it works perfectly, I’m done. If it has a bug or misunderstands a step, I don’t fix the code. I go back and fix my English instructions in the spec file and have the AI regenerate the code. This refines the source, not the output.

---

What This Changes for Designer-Run Agency Work

This shift from writing code to writing instructions fundamentally changes how agencies can operate, especially those run by designers or non-technical founders. It collapses the time and cost between an idea and a shipped product, turning technical execution into a commodity and elevating the value of clear direction and good taste.

Dimension	Old Way (Traditional Development)	New Way (Agent-Driven Development)
Skillset Required	Hiring specialized developers (frontend, backend, DevOps).	Writing clear, detailed instructions in plain English.
Speed to Prototype	Weeks or months. Requires scoping, hiring, coding, and testing cycles.	Hours or days. A single person can direct an agent to build a V1.
Cost to Build	High. Dominated by developer salaries and agency fees.	Low. Dominated by AI tool subscriptions and compute costs.
Iteration Loop	Slow. Changes require new tickets, developer time, and code reviews.	Fast. Refine the English spec and regenerate the code in minutes.

For my agency, this means we can now build and sell small, custom software solutions and automations that were previously unprofitable to scope. It’s a new, high-margin service offering that sits right next to design. It’s also a core part of my AI agents for builders framework.

---

My $0.02 — How I’d Roll This Out

Adopting agent-driven development isn’t about replacing your entire workflow overnight. It’s about finding one small, high-leverage point to start. My approach focuses on getting a real win in the first week to build momentum and prove the model’s value to myself and my team.

Day 1 — Find the pain. I’d look for the most boring, repetitive, and manual task in my own workflow. It could be compiling a report, sorting leads from a contact form, or checking websites for updates. The goal is to pick a task that is simple, rule-based, and happens frequently. Automating this one thing buys back real time, immediately.

Day 2 — Write the spec and build V1. I’d spend an hour writing a clear, one-page spec for an agent to perform that task. I would use the template from this post. Then, I’d feed that spec to Claude Code and generate the first version of the agent. I wouldn’t aim for perfection, just a working V1 that completes the core task.

Day 3 — Test and deploy internally. I’d run the agent on my own machine and for my own work. As I find edge cases or small bugs, I’d update my English spec—not the code—and regenerate. Once it’s reliable for me, I’ve got a proven asset. Now I can think about productizing it for clients or teaching it to my team, using my own experience as the case study. This is the heart of the talk-to-build stack.

---

FAQ

Do I need to be a coder to build an AI agent?
No, you need to be a good director. The primary skill is writing clear, specific, and unambiguous instructions in plain English. The AI translates your instructions into code. Your job is to be an excellent communicator, not a programmer.

What is the Cursor SDK?
The Cursor SDK is a toolkit that allows developers to create autonomous AI agents that can write, edit, and test code. It provides the framework for turning plain-language instructions into software that can interact with files, terminals, and other development tools to build applications.

Is this safe? Can an AI agent break things?
Yes, an agent can break things if given broad permissions. The key is to run agents in a controlled environment, especially during testing. Start with read-only tasks, review the code it generates before running it, and never give an untested agent direct access to production systems or sensitive data.

How much does it cost to run an AI agent?
The cost has two parts: building and running. Building uses tokens from AI services like Claude or OpenAI. Running the agent, if it’s simple, can often be done on a local machine or a small, inexpensive cloud server. For many internal tasks, the ongoing cost is negligible.

What is the difference between an AI agent and a chatbot?
A chatbot primarily engages in conversation to answer questions or provide information. An AI agent takes action. It can write files, execute commands, and interact with other software to autonomously complete a task without being guided through every step.

Can I use this for my agency’s client work?
Absolutely. Building small, custom AI agents to automate a client’s specific workflow is a massive value-add. It’s a new service you can sell that solves a direct business problem, moving you from a service provider to a solutions partner.

What are the limitations of current AI agents?
Agents excel at well-defined, self-contained tasks. They struggle with complex, multi-step projects that require broad context or subjective reasoning. They are also only as good as the instructions they are given. Vague instructions will produce vague or incorrect results.

---

Want help applying this?

Four ways to go deeper:

• Build with Builders. Join the Talk-to-Build community to learn to build AI-native websites, cinematic AI video, and agent-driven workflows you can sell.
• 1-on-1 working session. Book a screen-share with me — bring a real problem, leave with a working piece of it.
• Done-for-you. MK-Way builds AEO-ready websites, apps, and AI agent workflows.
• Quick question. DM me on Instagram or LinkedIn. I read every message.

---

Part of the AI Pulse series. If you commented “BUILD” on one of my videos — this is the breakdown.

Last updated: July 8, 2026.