Versus

one job, two designs

The second layer of the agent lineage: one question, one repo, one history, two ways to assemble them into the same 200K window. A twelve-turn clock walks both — the dumped side hits the ceiling at turn three and starts evicting material it will read again, the curated side compacts and keeps its headroom. Against the RAG tab: that one argues about fetching, this one about what is left.

One window, two ways to fill it same task, same 200K

one questionone repo · 2.1M tokens40 turns of history12 turns of work

they part company here — what gets assembled

Paste it all

a bigger window will fix it

everything▸concatenate▸send

system + tool defs8%
three PDFs, whole10%
the question2%
what the run read72%
left for the work8%

left for the work8%

evictions6

re-reads6

6 evictions in: the run is now paying to read files it already read.

Edit it

the window is the budget

everything▸select▸compress▸reserve▸send

system + tool defs8%
the standing rules2%
the passages that matter12%
history, as conclusions6%
the question2%
what the run read24%
left for the work46%

left for the work46%

compactions2

re-reads0

2 compactions: transcript dropped, conclusions kept, nothing re-read.

Both designs fill up; only one can afford to choose what it loses. A window that starts at 88% has no compaction strategy available to it, only eviction — and eviction is how a run ends up paying twice for the same file.

turn 12 / 12