AI agent
An AI agent is a language model placed in a loop where it can decide what to do next, call tools to do it, observe the result and continue until a goal is reached. The distinguishing feature is that the sequence of steps is chosen by the model, not fixed by the developer.
The parts
Four things have to be present. A goal stated in text. A set of tools — functions the model may call, described well enough that it can choose between them: search, read a file, run a query, send a request. A loop that feeds each tool result back in so the model can decide the next move. And a stopping condition, whether that is a success check, a step limit or a budget.
A common and more predictable variant is a pipeline: several specialised model calls in a fixed sequence, each with a narrow job. NorthernGo builds this way on the cloud path — an Architect that plans the structure, a Builder that writes the code, and a Critic that can attempt a repair of mechanical damage. Locally the Builder runs on your GPU and a syntax check blocks JavaScript that will not parse; there is no LLM rewrite. The order is fixed, which is exactly what makes the behaviour reproducible.
The distinction is worth keeping straight, because "agent" is used for both:
- Fixed pipeline — developer chooses the steps. Predictable, testable, limited.
- Autonomous loop — model chooses the steps. Flexible, capable of surprising you in both directions.
What to watch out for
A chatbot with a good system prompt is not an agent. If it cannot take an action in the world, it is a conversation. Most products labelled "AI agent" are one model call with a nice interface.
The real engineering problem is compounding error. A step that succeeds 95% of the time succeeds about 60% of the time over ten steps. Long autonomous chains fail in ways that are hard to see, because step four went subtly wrong and steps five through nine confidently built on it. This is also where a hallucination stops being a wrong sentence and becomes a wrong action.
Practical consequences:
- Keep chains short, verify between steps, and prefer a fixed sequence wherever the sequence is actually known.
- Cap everything. Step count, wall-clock time, token spend. Loops that call themselves are how you get an expensive surprise.
- Require confirmation for anything irreversible — deleting data, sending email, moving money.
- Log every tool call. Without a trace you cannot tell whether the model reasoned badly or the tool returned nonsense.
More agents is not better. Each additional autonomous hop multiplies the ways it can go wrong.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers; an agent acts. If the system can only produce text in reply to you, it is a chatbot no matter how it is marketed. If it can call functions, change state in another system and decide by itself which step comes next, it is an agent — and it needs the safeguards that implies.
Why do agents fail on long tasks?
Errors compound. Each step has a chance of going wrong, and later steps build on earlier ones without questioning them, so a small mistake early becomes a confident wrong result at the end. Shorter chains, verification between steps and a fixed sequence where possible are the standard mitigations.
Build it yourself
NorthernGo turns a plain-text description into a working web app with a database, login and a live URL. Local AI generation runs on your own GPU, is unlimited, and is free on every plan.