In plain English | GBS SE Agent

Download the guide

GBS SE Agent, explained in plain English (PDF) covers everything on this page in more depth: everyday comparisons, pictures, where the brain can live and what that costs, how it compares to the better-known tools, and what comes next.

What it is

Imagine someone on their first day. They do not know where anything is, but they are quick, never get bored, and have three good habits:

They look before they touch. They do not guess where the login code lives; they go and find it.
They ask before anything risky. Running commands, changing settings, opening a file that looks like it holds a password: they stop and check first.
They show their receipts. When they finish, they tell you exactly which files they touched, so you can look before anything goes out.

That is the whole product. Everything technical in these docs exists to force those three habits onto software that would otherwise happily guess.

Six steps in a row: you describe the job in your own words; it looks, searching and reading the files; it asks before anything risky; it changes the code; it checks, running the tests and reading errors; and it shows you every file it changed. An arrow runs from the checking step back to the looking step: if a check fails, it goes back and fixes it, as many times as it takes. — Steps two to five repeat until the work is actually done. That loop is what makes it an agent rather than something that answers and stops.

Steps two to five repeat until the work is actually done. That loop is what makes it an agent rather than something that answers and stops.

Three words you will keep hearing

The model is the brain. The agent has no intelligence of its own; it is the hands, the eyes and the rulebook. The thinking is done by a model, which can run on your laptop, on a company server, or be rented over the internet.
Tokens are pieces of words, about four characters each. 750 words is roughly 1,000 tokens. You pay for tokens, and a model can only hold so many at once.
The context window is the desk. Everything the agent is thinking about has to fit on it: the house rules, your request, every file it read, every test result. When the desk fills up, the agent tidies it, throwing away old output first and summarising only as a last resort.

Where it helps, and where a person is still needed

It is good at	A person is still needed for
Explaining a codebase nobody on the team wrote	Deciding what should be built at all
Repetitive changes across many files	Judging whether the result is actually right
Writing the first version of tests	Anything where being confidently wrong is expensive
Narrowing down where a bug lives	Context nobody wrote down
Following conventions it was taught once	The last review before it ships

The one-sentence version

An assistant that can read and change a codebase on its own, shows you every step it takes, asks permission before anything it cannot undo, and can run entirely on your own hardware so client code never leaves the building.