What an AI agent can and cannot do with your files

The difference is the file system, not the model

A chat assistant and an agent can be the same model. What separates them is access: the agent is handed a working directory and a set of actions - read a file, list a folder, run a command, write a result - and it decides which to use. Nothing about the reasoning changes. What changes is that the reasoning is now grounded in the actual contents of your project rather than in your summary of it.

That is why the interesting question is never "is the model smart enough". It is "what did it read, what did it run, and what did it change". Those three are answerable. Whether a model understood you is not.

What this arrangement is good at

• Reading a lot of material that you would otherwise skim. A folder of PDFs, a set of specifications, a year of correspondence: an agent can go through all of it and come back with the three places that contradict each other.
• Extracting structure from documents that have it but do not expose it - pulling a sheet index out of a drawing set, or a schedule out of a scanned table.
• Repetitive transformation with a rule you can state. Renaming, reformatting, reconciling two lists, turning notes into a table.
• Writing the first version of something formulaic - a transmittal, a meeting record, a covering note - from material that already exists.
• Building a small tool. You describe the calculation you keep doing by hand, and you get something you can run, read, and correct.

The four ways it goes wrong

1. It answers confidently from the wrong file

An agent that cannot find the current revision will often use an old one rather than say it could not find anything. The output is fluent and internally consistent and describes a superseded document. This is the most common failure and the hardest to notice, because nothing about the answer looks wrong.

The defence is structural: keep one obvious current folder, and ask what the agent read before you ask whether it is right.

2. It fills a gap instead of naming it

Asked for a value the source does not contain, a model will frequently produce a plausible one. In prose that reads as a reasonable assumption. In a table of extracted quantities it reads as data. The two are indistinguishable once the answer has been copied somewhere else, which is the argument for keeping extraction output next to the document it came from rather than in a fresh spreadsheet.

3. It is confident about arithmetic it did in its head

A language model computing a number by prediction is not calculating. Anything with a formula in it should be run - in a sheet, in a tool, in a script - rather than asserted in a sentence. An agent that can execute code is much safer here than one that can only write, and it is worth insisting on the executed version.

4. It changes more than you asked

Write access is the whole point and also the whole risk. An agent that reformats a file while fixing one line has done something you did not ask for and may not spot. Version control, or at minimum a copy, is the answer, and it is the same answer as for any other tool that edits your work.

Review is the part that stays yours

None of this moves professional responsibility. An agent output is a draft: something that saves the first hour, not something that replaces the review. That is not a limitation of the current generation of models that a better one will remove - it is what it means to be the person who signs.

The practical consequence is that the review has to be cheap, and review is cheap when the output sits beside its source. A quantity next to the drawing it was read from can be checked in seconds. The same quantity pasted into a spreadsheet three folders away cannot be checked at all without redoing the work.

Where CoLateral fits

CoLateral is a workspace built around exactly that arrangement: an infinite canvas holding your files, the agents that read them, a terminal, notes and sheets, and the tools you build yourself. The agents work on the folder on your own disk - the project stays where it is, in the format it is already in.

Two of the shipped tools exist because the failure modes above are worth engineering against directly: drawing intake reads a set and shows you what it found next to the page it found it on, and the PDF sheet indexer builds an index from the sheets themselves rather than from what someone typed into a transmittal.

What it is not: there is no cloud project storage, no multi-user collaboration, and no CAD or analysis kernel underneath any of it. It is a workspace for your files and the agents you point at them.