What is an AI agent
Last updated: 20 September 2026
An AI agent is software that acts on your behalf. You give it a goal, and instead of answering you it goes and does the work — reads files, runs commands, calls APIs, opens web pages — then looks at what happened and decides the next move.
A chatbot can describe a fix beautifully, but it stops at the description. Asked to change a config file, it tells you what to change; you copy, paste, save, restart. An agent changes it and tells you it's done.
The difference is a loop
A single exchange looks like this:
your question → model → answer → doneAn agent looks like this:
your goal → model picks a step → execute → read result → model picks next step → …
↓
goal reached / stuckThat "read result" is the whole trick. When an agent's instruction runs, it produces real feedback — a command fails, a file isn't there, a page returns 404. That feedback goes back to the model and shapes the next instruction. Which is why it can correct itself: wrong path, try another; request timed out, retry.
This loop is what people mean by tool calling. The model isn't restricted to conversation; it has a set of functions, and each one maps to a real action.
A concrete example
Take "upgrade this project's dependencies to the latest versions":
| Step | Chatbot | Agent |
|---|---|---|
| Look | Tells you to run npm outdated | Runs it, sees 12 packages behind |
| Edit | Hands you a package.json snippet | Edits the file, installs |
| Handle errors | You paste the error back, it revises | Sees the conflict, rolls one package back, retries |
| Verify | Reminds you to run the tests | Runs them, two fail, goes back and fixes |
The gap isn't intelligence. It's whether the thing has hands. Same underlying model, one talks and one acts.
It gets things wrong, by design
Agents fail differently from chatbots. When a chatbot is wrong, you notice while reading. When an agent is wrong, a file is actually deleted, a branch is actually pushed, an API call actually went out.
Three patterns cover most of it:
- Right direction, too much force — you say "clean up unused files" and it removes more than you pictured
- Reading failure as success — a command exits non-zero and it moves on without checking
- Long runs on a wrong premise — it misread the goal at the start and every step since builds on that, so by the time you look, ten minutes are gone
This is why every serious agent tool ships some kind of stop-and-ask mechanism. Approval gates, sandboxes, plan mode — different names, same purpose: let a human look before an action that can't be undone.
Why this appeared now
The idea is old; earlier attempts existed. Two things had to land together for it to work: models that reliably emit structured instructions, and context windows large enough to hold a whole task's back-and-forth. Without either, the loop falls apart within a few turns.
Getting started
Most agents are command-line programs — Claude Code, Codex CLI and similar. You install one and talk to it in a terminal. They're independent, so running several means running several terminal windows.
Aiglade handles that layer: it collects those command-line agents into one desktop window with a shared task queue and a single approval gate. The agents stay what they were; you just stop switching windows.