Guiding your AI agent like a new hire
1. It is not a genius, it is a newcomer
The right picture is not "a machine that knows everything". It is "a new hire who does not know the company yet". Brilliant on generalities. Ignorant of your setup. A company that welcomes a new arrival well does two things. It gives the access it needs. It shows examples. One that hosts badly does the opposite: it hopes the newcomer will guess.
So the first guiding move is the start. Launch your agent. Hand over access to your setup step by step. Not all at once, until it drowns. Not at all, until it spins in circles. The access that matters — credentials, paths, repositories — lives in a safe place it can consult. A vault. No need to repeat them afterwards. It reads them, like a hire reads the badge instructions.
2. Brick by brick, never in bulk
Nobody joins a company and understands everything in a day. You hand over one brick first. Very small. You let it build it. You watch. Then you point it to the next. The opposite trap — describing all ten bricks at once to "save time" — produces exactly the reverse. It tangles them. It invents the order. Everything must be redone.
At each brick, ask it discovery questions, not orders. Instead of "write this script", ask: "Can you see a file that is supposed to do this?" It digs. It finds. It may be wrong — you point it. That is how it learns where things live in your project. Not in theory.
3. A search function, built to be kept up to date
Very quickly, the bottleneck stops being doing. It starts being finding. After a few bricks, your agent no longer knows where it stored what. The answer is not to feed it the information again and again. It is to have it build, from the start, a search function. A small tool that says where things are in your setup.
This is the crux: ask it to specify it, then to code it, then, every time you discover a new place, update that function. It searched, you pointed, it found where the configuration lives. That discovery goes straight into the search function. Next time, no more searching. It asks. The tool grows with the company. It does not go stale.
4. Show it once, record it always
This is the hardest rule, and the one most often forgotten. If you never show a move once, and you do not record it where the agent looks for its next tasks, you will struggle to work side by side with it. Not because it is slow. Because you will miss each other. Each time, it repeats the same mistake. You, the same correction.
The right rhythm: I show, it understands, I write it where it will look next time. I check it found it before moving on. The note is not an end-of -session summary meant to be forgotten. It is a working tool meant to be re-read.
5. Talk to it like a colleague you are training
Let us sum up the attitude that works: talk to it as if you were guiding a new hire around the company. Reassure. Point. Give context. Show an example. Do not let it guess implicit rules. In a company, that kind of unwritten rule costs a newcomer weeks. With you, it costs a whole session every time it is left unsaid.
Above all: do not keep the discovery to yourself. What you showed once becomes, once recorded, a reflex. The goal is not that the agent gets smarter. It is that your way of working is somewhere. Reusable. Independent of your presence and your memory. That is exactly the difference between delegating once and building a team.