AI Agents: What They Are, How to Build One & Where They Fail

👁 139 reads

The word “agent” has been slapped onto everything with AI in the name, and it’s lost all meaning. But the difference between a chatbot and an agent is concrete and can be summed up in one sentence: a chatbot answers, an agent acts. The rest of this article unpacks that sentence and shows you how to build your own agent without writing code.

How AI Agents Differ from Chatbots

A regular chatbot receives a question and returns text. Nothing else happens. If you ask it to send an email, it will write the email content but won’t send it.

An agent receives a goal, not a question, and has tools at its disposal. It can check something online, read a file, write to a database, send a message. It decides what order to do things in and checks the results along the way.

That’s the entire difference. You tell a chatbot “write a summary of this file.” You tell an agent “every morning, check for new files in this folder, summarize each one, and send me a single message with everything.”

What an AI Agent Is Made Of

Regardless of what platform you build it in, an agent has four parts.

1. Model. This decides what to do in the next step.
2. Tools. Everything the agent can use to accomplish something beyond writing: search engine, file access, email sending, database queries.
3. Memory. A record of what it’s already done and what it’s learned, so it doesn’t start from scratch at every step.
4. Trigger. What starts it running: a time, a new email, a new file, a click.

If any of these four is missing, it’s not an agent - it’s a chatbot with prettier packaging. Most often, tools are what’s missing.

How to Build an AI Agent Without Code

The simplest path goes through automation tools where you connect ready-made blocks. The pattern is always the same.

Step one: start with a single task you do manually. Not a vision of an assistant that runs your company. One activity you repeat every Monday. This is the only advice on this list that truly determines success.

Step two: break it down into steps as detailed as if you were explaining it to a new person. Where to look, what to check, what to do when something’s missing. That last point is what everyone forgets and what causes agents to get stuck.

Step three: give the agent only the tools it needs. Not “access to the inbox,” but “read messages from one label.” The narrower the permissions, the fewer things can go wrong.

Step four: run it in dry-run mode first. Let it write out what it would do for a week instead of actually doing it. You’ll see all the places where it made mistakes you didn’t expect.

Step five: only now turn on actual actions, and only on something reversible.

Where AI Agents Actually Break

We’re writing this from experience with our own automations that run daily.

On atypical situations. An agent works great as long as the data looks like it did during testing. The first file in a different format can throw it off and, worse, the agent usually doesn’t report the problem - it just does something that looks sensible but is wrong.

On silent failures. A model can return an empty response with a success status. We measured this: in our test of ten models, three of them in one task returned status 200 with empty content, without any error message. If your agent doesn’t check for this, it will save nothing and move on.

On costs. An agent that thinks out loud pays for every step. With models that have mandatory reasoning, thinking tokens make up as much as 96 percent of all output tokens, and output is the more expensive side of the pricing. The bill for an agent running around the clock can be surprising.

On loops. An agent without a step limit can spin in circles. Always set a hard limit on attempts and a hard limit on cost.

What AI Agents Are Actually Good For

For repetitive, boring tasks. Browsing, organizing, summarizing, converting from one format to another. Here an agent pays for itself from the first week.

For monitoring. Checking if something has changed and only speaking up when it has. This is probably the best first project for anyone.

Not good for decisions you can’t verify. If you can’t assess in five seconds whether the agent did well, don’t give it that task. Not because it can’t handle it, but because you won’t notice when it doesn’t handle it.

Which Model to Choose for an AI Agent

For an agent that runs in a loop, you choose differently than for a one-time conversation, because cost multiplies by the number of steps. We measured ten models on writing tasks in Polish (a heavily inflected language, useful as a proxy for non-English performance) and the difference in cost between models with identical accuracy reached thirty-two times.

On top of that, there’s something few people remember: the price visible on a model’s card doesn’t have to be the price you pay, because the same model is hosted by a dozen companies with their own rates. We broke this down with data from fifteen models.

How Much Does a Writing AI Agent Cost

With an agent running in a loop, cost multiplies by the number of steps, so model choice becomes a financial decision. We broke this down with concrete numbers: how much an AI-written article costs, from four hundredths of a cent to nineteen cents for the same text.

Running a Model on Your Own Computer

An agent doesn’t have to run in the cloud. If you care about keeping data on your own machine, we tested how this works in practice: a model run locally.

Where Automation Ends

When planning agents, it’s useful to know which activities can be handed off to a tool today and which can’t. We broke this down with specifics in the article about AI and work.

Before You Choose a Model for Your Agent

When working in a loop, cost multiplies by the number of steps, so it’s worth knowing that seven out of eighteen models no longer allow you to disable reasoning. If your code sends that parameter as a fixed value, it will stop working entirely after you change models.

// read moreBuild an AI Agent in Python Step by Step (76 Lines)GuidesGumloop Tutorial: Build Your First AI Agent Without CodeGuidesAI Agents from Scratch: How They Work & How to Build OneGuides

Kinetyra