Guide · August 2026

A chatbot answers.
An agent executes, checks, and tries again.

The word "agent" now gets used for almost anything with AI inside, which is why it has stopped meaning anything. This guide uses an operational definition — one you can check by looking at the tool — and uses it to walk through what agents actually do today, where they fail, and which ones you can use.

No hype · failure modes included

The definition that works

An agent is the one that picks the steps

Academic definitions talk about perception, environment and autonomy, and help you decide nothing. There is a more useful one because it is checkable: an agent is a system that, given a goal, decides for itself what steps to take, executes them and checks the result.

A chatbot hands you text. You ask how to deploy your app and it explains how. Execution is still yours, and the text can be correct and still useless if something in your environment does not match.

An automation executes steps, but you wrote those steps beforehand: a Zapier or n8n flow does exactly what you drew, every time. That is its virtue and its limit — when reality deviates from the drawing, it fails.

An agent executes steps you never gave it. You state the goal, it looks at the real state of things, chooses, acts, sees what happened and corrects. That is where its value is, and its risk: for the same reason it adapts, it is not reproducible, which rules it out for processes that must come out identical every time.

The three-second test: if you can write out the complete list of steps it will take in advance, it is not an agent — it is an automation with a language model inside. Still useful, but not the same thing, and it does not fail the same way.

The landscape

Six tools and what each is for

Ordered by kind of work, not by popularity. The open source ones are marked as such; none of these are affiliate links.

Coding agent · terminal

Claude Code, Aider

Agents that work over your repository from the command line: read the project, edit files, run tests. Aider is open source; Claude Code is proprietary and comes with the Anthropic subscription.

Cloud agent · app building

Manus

A cloud service built so someone non-technical can describe an idea and receive a deployed app. Everything happens on their infrastructure, and it is paid with burnable credits.

Frameworks · build your own

CrewAI, LangGraph

Python libraries for assembling your own multi-agent system. They are not installed and used: they are programmed. Open source, with the cost that you maintain an internal project.

Automation with AI nodes

n8n, Zapier

Workflow tools that have added AI nodes. They remain deterministic — you draw the steps — which is exactly why they are the right answer when the process is stable and high volume.

Local orchestrator · desktop

Lienox

A desktop app that runs several agents on your machine with your own AI keys, with an agent browser, a board and scheduled runs. Proprietary, with a permanent Free plan. It is ours: bear that in mind while reading the rest.

Local models

Ollama, llama.cpp

The layer underneath: they serve open-weight models on your machine. Not agents, but what you need if the requirement is that no data leaves your network.

The differences, laid out

Chatbot, automation and agent compared

Most disappointment with "agents" comes from picking the wrong column for the problem you actually had.

Aspect Chatbot / Automation Agent
Who picks the steps You, before it runs The system, while it runs
What happens when something changes It breaks, and has to be edited It sees it and adapts — or fails in a new way
Reproducibility Identical on every run Variable: two runs can differ
Cost per run Predictable Variable, and rises with task length
Work it suits High volume, clear rules, zero exceptions Cases with judgement, exceptions and shifting context
How you supervise it A step-by-step execution log You have to review the result, not just the log
Main risk Breaking silently when an API changes Doing something plausible and wrong, confidently

What they actually do today

Five tasks, with the failure mode of each

1

Writing and fixing code in a real project

By far the most mature use. It works well in projects with tests and clear structure. It goes wrong in large repos without tests: the agent makes a plausible change, nothing contradicts it, and the failure shows up three days later.

2

Researching and summarising with sources

Searching, opening, reading and synthesising is among the things it does best today. The typical failure is not inventing the source but over-trusting the first one it finds: tell it explicitly to cross-check and to say when it found nothing solid.

3

Operating web interfaces with no API

An agent with a browser fills forms, pulls data out of dashboards and runs checks no API exposes. It fails on captchas, on two-factor prompts, and on interfaces that depend on precise drag and drop.

4

Watching and reporting

Checking every morning whether something changed and saying it in two lines is an unglamorous use and one of the most profitable. The usual failure is over-reporting: you have to tell it explicitly to stay quiet when there is nothing to say.

5

Preparing work for you to decide on

The draft, the comparison, the list of things to review. This is the pattern with the best value-to-risk ratio, because a mistake never reaches the world without a person seeing it. If you are starting out, start here.

How to ask an agent for something

The instruction matters more than the tool

The difference between an agent that helps and one that disappoints is almost never the model: it is how the work was handed over. Three rules cover most of it.

State the goal, not the steps. If you dictate the steps you are using an expensive automation. The value shows up when you give it the result you want and the criteria for judging it.

Tell it what to do when it does not know. The most expensive agent failure is inventing something plausible rather than admitting it found nothing. One sentence — "if you find nothing solid, say so and stop" — removes a large share of the problem.

# Bad: dictating steps (this is an expensive automation) Open X, copy column B, paste it in Y, hit save. # Good: goal + criteria + what to do on failure Go through yesterday's orders and tell me which ones look off and why. If something does not add up, do NOT fix it: flag it and move on. If nothing is odd, reply only "all normal".

Frequently asked questions

AI agents

What is an AI agent?

A system that, given a goal, decides for itself what steps to take, executes them and checks the result. The difference from a chatbot is that it executes rather than only answering; the difference from an automation is that the steps were not written in advance. That second difference is the one that matters when picking a tool.

How is an AI agent different from a chatbot?

A chatbot returns text and execution stays with you. An agent acts on the world: it opens files, runs commands, browses, writes. The practical consequence is that an agent can be wrong by doing, not just by saying, which is why it is worth starting with tasks where a person reviews the mistake before it ships.

What are AI agents actually good for today?

Solid results: writing and fixing code in projects with tests, researching and summarising with sources, operating web interfaces with no API, watching things and reporting, and preparing drafts for a person to decide on. Uneven results: long unsupervised processes, anything gated by captcha or two-factor, and tasks that must come out identical every time.

Do AI agents replace automations like Zapier?

No, and believing they do is an expensive mistake in both directions. For a stable, high-volume process with no exceptions, a deterministic automation is cheaper, faster and more auditable. Agents win exactly where those tools struggle: exceptions, judgement, and tools with no API. Most people end up using both.

What is the best AI agent?

It depends on the work, and anyone who answers otherwise is selling. For code in a repository, a terminal agent like Claude Code or Aider. For building an app without touching code, a cloud service like Manus. For automating several fronts from your own machine with your own keys, a local orchestrator like Lienox. For assembling a bespoke system, CrewAI or LangGraph.

Are AI agents safe?

The real risk is not that they rebel: it is that they do something plausible and wrong with great confidence, and that they hold enough permissions for it to matter. The three measures that actually change the outcome are minimal permissions, starting with tasks whose result a person reviews, and running in an environment where you can see and undo what they did.

How much does using an AI agent cost?

There are two bills and it helps to look at them separately. The platform — subscription, credits, or nothing if it is open source — and model inference, which on agentic tasks climbs faster than people expect because each step re-reads the accumulated context. Flat pricing protects you from the first; from the second, only choosing the model well and scoping the task does.

Try an agent on a task that already annoys you

Free forever on Linux, Windows and macOS. Runs on your machine, with your own AI keys.

No credit card · BYOK