Main idea

The reasonable starting point for handing building-management emails over to an agent comes down to two gestures: create a context.md, and give two or three instructions in the conversation — aggregate, keep a history when there are several stages, write it all in there. That setup has no chance of holding: it will lose the history, it will swell, it will produce duplicates on names.

That isn't a flaw, it is its function. What it produces that lasts isn't the file — it is what you see by watching it work: how the agent groups the material when you don't tell it how, where it overwrites what you wanted to keep, which ambiguities in your domain trip it up. None of those observations is available before running it on your own material, and all of them are exactly what you need to know in order to write the next system's directives. The first system is the instrument that reveals them.

It follows that a first version thrown away isn't a failure, and that a second version started with the intention of throwing it away may well survive. What carries over from one version to the next isn't the structure, it is the knowledge of what breaks — and that knowledge can't be read in any description of a finished system, because a finished system shows only its answers, never the breakdowns that motivated them.

Why it matters

This lifts the most common blockage at the start: looking for the right architecture before having anything to judge it by. The first attempt doesn't have to be defensible; it has to run on real material long enough to produce observations.

It also changes what you watch during that period. You don't assess the quality of the resulting context, you record the gaps between what the agent did and what you expected — it is those gaps, not the file, that will be reused.

Nuances and limits

The learning is about the structure, not the material. The five hundred emails ingested during that period aren't disposable along with it, and a botched first version can cost a data migration that exceeds what it taught.

And the observation is only worth something if the material is your own. A trial run on a sample dataset reveals neither your contractors' shared surnames nor the thresholds that matter in your world: it teaches you how the tool works, not where your domain breaks.

Open questions

• How long must a first system run before what it shows is representative rather than anecdotal?