How to create an AI agent

To create an AI agent, pick one job with a clear finish line, write its rules down, connect it to your systems with read access first, test it on last month's real cases, then run it beside a person before it acts alone. The tool you build it in is step 5 of 7, not step 1.

I'm Hayat Amin. I spent twenty years as a technology chief financial officer, sold three companies in that seat, and now build agents inside companies myself as a forward deployed engineer. Most of the agents I see fail weren't badly coded. They were pointed at a job nobody had written down. Below are the 7 steps I follow, worked through one example, then the 6 places Google suggests you build one, read from each vendor's own pages on 13 October 2026.

The example: a maintenance request agent

Say you run a property management company in New York. A tenant texts at 11pm that the kitchen sink is leaking. Today someone reads that in the morning, opens the property system, writes a work order, decides how urgent it is, picks a plumber and texts the tenant back. That's 6 jobs, 4 systems and one person who can't be off sick. It's a good first agent because the finish line is plain: a correct work order with the right vendor on it. I'll use it through every step.

7 steps to create an AI agent

1. Pick one job with a finish line

An agent needs a job you could hand to a new hire on day one with a checklist. "Handle maintenance" is a department. "Turn every inbound repair message into a work order with an urgency and a vendor" is a job. If you can't say what done looks like in one sentence, you can't test it, and step 6 will never end.

2. Write down the rules your best person follows

OpenAI's practical guide to building agents says an agent has 3 parts: a model, tools, and instructions, the explicit guidelines and guardrails for how it behaves. The instructions are the part companies skip. Sit with the person who does the job well and write down what they do. A leak under a sink is urgent, a dripping tap isn't. No heat in January is an emergency. Never send a vendor over a set amount without a manager's yes. That document is the agent. The model only reads it.

3. Decide whether it needs to be an agent at all

Anthropic's engineering team draws the line in their post on building effective agents. A workflow is a system where models and tools run through predefined code paths. An agent is one where the model directs its own process and decides which tools to use. Their advice is to find the simplest solution possible and add complexity only when you need it. I agree with them. Writing the work order is a workflow. Troubleshooting with a tenant who isn't sure which pipe is leaking needs an agent, because the next question depends on the last answer.

4. Connect its tools, read access first

This is where most of the work is, and it's the part I'd spend your budget on. The agent has to read the tenant's message, look up the unit, check the vendor list and write the work order. The Model Context Protocol, MCP, is an open source standard for connecting AI applications to outside systems. Your property system has to let it in. Buildium includes its Open API only on its Premium plan, from $400 a month, against $62 on Essential. Rent Manager offers a web services connection with both read and write access to your database. Give the agent read access for the first weeks. Write access comes after it has earned it.

5. Choose where to build it

Pick the place your team can maintain, not the one with the best demo. The 6 options are compared in the next section. If you're on AppFolio, check first whether you need to build anything. Its Realm-X Maintenance Performer already troubleshoots with residents, writes the work order, classifies the urgency and routes the vendor, under human monitoring and supervision. Buying that beats building it. I wrote about where that line sits in should my company build or buy its AI agents.

6. Test it on last month's real cases

Pull 50 real requests from last month, with the work orders your team actually wrote. Run the agent on them and compare. Count where it picked a different urgency or a different vendor, and read why. Every miss is either a rule you forgot to write in step 2 or a system it couldn't see in step 4. Fix those and run the same 50 again. Don't move on until the misses are ones your best person would also argue about.

7. Run it in shadow, then give a named person its exceptions

For 2 to 4 weeks the agent drafts every work order and a person approves or corrects it. Then it sends the routine ones alone and passes anything it isn't sure of to a queue. Put one name on that queue. An agent with nobody owning its exceptions turns into a second inbox nobody reads. I covered the controls I put around that in how do I make AI agents safe enough to run real operations.

Where to create an AI agent: 6 places compared

These are the 6 that Google completes when you type the question, in the order it suggests them, with free and from scratch at the end. I read each vendor's own documentation and pricing page on 13 October 2026. Prices are only where the vendor prints them.

1. Microsoft Copilot Studio

You describe what the agent should do in up to 1,024 characters, and the AI drafts its name, description and instructions, then suggests triggers, knowledge sources and tools. You add tools, test it and publish. It's good for a company that already lives in Microsoft 365 and wants the agent in Teams. It's wrong for anyone who wants to try it on the trial, because a trial licence can't publish. Microsoft prints Copilot Credit packs at $200.00 a month for 25,000 credits and says an Azure subscription is required to use agents.

2. Claude

The Claude Agent SDK gives you the same tools, agent loop and context management that run Claude Code, in Python and TypeScript. It connects to your systems over MCP. It's good for an agent that runs on a schedule and does real work across several systems, which is most back office work. It's wrong for a team with nobody who can read code. I set out how I run it day to day in how do I use Claude agents to run business operations.

3. ChatGPT

Check the date before you start here. OpenAI told developers on 3 June 2026 that Agent Builder, its visual canvas for multi-step agents, is deprecated, and it's scheduled to shut down on 30 November 2026. OpenAI points builders to the Agents SDK or ChatGPT Workspace Agents instead. It's good for a company already on a ChatGPT business workspace. It's wrong for anyone who builds on Agent Builder this month and has to move it again in 7 weeks.

4. Gemini

On a personal account you open Gems, choose New Gem, give it a name and instructions and add files. That's a helper with standing instructions, and it can't act in your property system. Google says that from November 2026, Gems on personal accounts go away and turn into skills. For a company, Gemini Enterprise has a Workflow Builder on its Standard, Plus, Pay-as-you-go and Frontline editions, where you describe the agent in a prompt or build it in designer mode. It's good for a team on Google Workspace. A Gem is wrong for anything that has to write to another system.

5. n8n, for free

n8n's Community Edition is the free edition, the one it runs without a licence key, and its AI Agent node lets you build an agent that needs at least one tool connected. It's good for the step 3 workflow, because you can see every path drawn on a canvas. It's wrong if you can't run a server, and the free edition leaves out SSO, sharing and Git version control. Free means free software. You still pay for the model and the hosting.

6. From scratch, with a framework

OpenAI's Agents SDK calls itself a lightweight package with very few abstractions. LangGraph calls itself a low-level runtime for stateful, long running agents that mix hand-coded logic with model driven steps. Both are good for an engineer who wants full control of the loop. Both are wrong as a first project for a company with no engineer, because you're signing up to maintain the code.

What I'd do this week

Pick the one job in your company that eats a named person's Monday. Write its rules on one page. Then check whether your existing software already ships the agent before you build one. If it doesn't, the build is mostly step 4, joining the systems, and that's the work I do through Beyond Elevation, my firm. I sit inside the company, connect the systems the agent needs, write its rules with the person who knows them and stay until it runs without me. The scope is at meethayat.com/services/fde.

If you want a second pair of eyes on which job to give your first agent, I do a free audit call at beyondelevation.com/call/hayat.

About Hayat Amin

I'm Hayat Amin, and I've spent twenty years in technology, most of them as chief financial officer of companies growing faster than their systems. I sold three of those companies in that seat, with American Express and TripAdvisor among the buyers, and carried three FT 100 fastest growing listings along the way.

I'm exceptional at creating agents that survive contact with a real company, because the finance seat taught me where the week breaks before anyone writes code. I connect the systems that were never built to talk to each other, write the rules down with the people who know them, and put a live dashboard in front of the chief executive instead of a month end pack. I also value and monetise the intellectual property and data a company already owns, and sit beside founders from the first conversation to the wire transfer on an exit.

I'm available now for fractional chief financial officer work and AI operations work through Beyond Elevation.

Questions people actually ask

Where these facts come from

Every product fact and price on this page is from the vendor's own documentation or pricing page, read on 13 October 2026. The 6 places are the completions Google suggested in the US that morning for how to create an AI agent. The property management company is an example, not a client, and the 50 cases and 2 to 4 weeks are how I run tests, not a published benchmark. OpenAI's help pages on custom GPTs refused my request, so I haven't said anything about which plans can make them.

The sources named on this page:

• A practical guide to building agents
• Building effective agents (19 December 2024)
• What is the Model Context Protocol (MCP)?
• Agent SDK overview
• Quickstart: Create and deploy an agent
• Microsoft Copilot Studio pricing
• Agent Builder guide
• Deprecations
• OpenAI Agents SDK
• Use Gems in Gemini Apps
• Create an agent in Gemini Enterprise
• Community Edition features
• AI Agent node
• LangGraph overview
• Realm-X Performers demo
• Buildium pricing
• Rent Manager software customization

Vendors change plans and retire products quickly in this market, so if you're reading this months later, open the links before you trust a price.