How an agent turns a site visit into project memory

After a first site visit, you usually do not come back with a project. You come back with a folder full of the site's real life.

Phone photos. A few angles of the house. The fence. The driveway. The utility shed. Trees. Damp spots. Places that are still just soil after construction. One photo from a corner nobody will remember later. Measurements. A rough sketch. The owner's questionnaire. Wishes. Doubts. A passing comment near the fence: "This could be a good place to sit."

All of that is easy to lose. Not physically. The files are still there. What gets lost is the connection between them.

One photo shows an old foundation. In the owner's brief, that same area is mentioned as a possible lounge zone. The measurements show a narrow passage. The first AI visual check already puts a nice terrace there, but the model does not know that the place is damp and awkward to reach from the house.

That is why we started using agents that work inside the project folder. Not instead of a designer. Not instead of a site visit. Not instead of proper measurements. As working memory for the project.

We start with the project folder

When a new site project starts, we do not begin with pretty images. First we put the material into a structure.

We keep this structure as a folder template that can be copied for each new project. It is still a first working version, and we keep improving it.

This article is about the start of the project, so only a few stages matter here. The folder holds the raw material from the visit: photos, video, measurement notes. A measured plan. The client brief and site analysis. The first AI visual checks.

Photos stop being just photos

Raw site photos are useful only while someone still remembers what is happening in them. After a week, one angle starts to look like another.

So the agent's first job is not to imagine. It is to look. We use models that can work with images. They can go through the photos carefully and help build a first description: what is visible in each shot, which objects repeat across photos, where the house and structures are, where there are trees and problem spots.

This is not a final inventory. It is the first index of the site.

The value is in comparing layers

The useful part is not that the agent describes photos. The useful part starts when it compares the layers of the project.

Photos say one thing. Measurements add another. The owner brief adds a third. The first AI visual checks suggest a fourth.

For example, the photos show trees already growing along the fence. The owner says they want more privacy. A visual check adds a dense green screen along the same boundary. At first glance, it all lines up. Then the photos show that the passage is already narrow, and the sketch says technical access is needed there.

Why we do not ask the agent to design the site

Because it will. And that is the risk.

If you give an agent a folder of photos and ask it to "propose a landscape solution", it can produce a confident answer. But confident language does not mean the decision has been checked.

It may misread scale. It may treat temporary construction clutter as a permanent object. It may miss an important tree. It may suggest planting where you need machinery access.

So we do not launch the agent as the "site designer". We use it as an archivist, analyst and careful assistant.

The first AI images also become part of the memory

Early AI visual checks should not live in a separate universe. If images are in one folder, site photos in another, measurements in a third, and the owner's brief somewhere in a chat, confusion starts quickly.

That is why we put first visual checks into the project folder and ask the agent not just whether they are "nice", but what they keep, break or reveal.

What should exist after the first pass

After the agent has done this first pass, the project folder no longer feels like a pile of disconnected files. It has memory.

The agent can prepare: a photo index, a list of visible objects and zones, a first site description, connections between photos and wishes, a list of contradictions, questions for the owner, risks in the first visual ideas.

The point

AI is useful here not because it is "smart". At the start of a project, its value is simpler: it helps keep context from falling apart.

A site is not one file and not one image. It is a set of facts, observations, wishes, constraints and hypotheses. If they are not connected early, the team keeps rebuilding the picture in their heads.

An agent inside the project folder makes that picture external. It does not decide for the designer. It helps the team see what they are actually working with.