Unit 04.02: Context that grows until it stops working

A transcript grows every turn and is re-sent every turn, which makes the cost of a long run rise with the square of its length.

Quadratic, not linear

Twelve turns of a conversation, tracked as a growing transcript against a fixed set of fields.

The table below compares both.

 turn   transcript   fields
    1          340      180
    4        1,360      180
    8        2,720      180
   12        4,080      180

after 12 turns: transcript 4,080 chars, fields 180 chars
the transcript is re-sent every turn, so total cost is quadratic

The arithmetic is the point. Each turn adds to the transcript *and* the whole transcript is sent again, so turn twelve costs twelve times what turn one did - while a fixed set of fields costs the same every turn.

That is why long agent runs get expensive in a way that surprises people. The per-step cost looks stable in testing, where runs are short, and rises steeply in production, where they are not.

The mistake this prevents

The mistake is measuring cost per step on short test runs. A ten-step average tells you very little about a forty-step run if the context accumulates, and the difference is a factor of four rather than a few percent.

Takeaway

Extract facts into fields rather than carrying a transcript. Accumulating context makes long runs quadratically expensive, and test runs are too short to show it.

Mark complete