Context Budgeting

Context budgeting is the fixed allocation of token amounts to each stage of an AI agent pipeline. Instead of copying the entire project history into every model call, each step receives a defined context budget: headings instead of full texts, numbered summary blocks instead of raw transcripts, targeted reloading instead of constant repetition. This keeps cost, latency, and memory per task calculable, because the token count per step grows with the number of steps, not with the length of the project. Typical building blocks are a compact task card, a numbered fact list, and a short session log that is carried forward and merged into the next step. Context budgeting is therefore the operational foundation of economical agent systems — in contrast to the classic full-context approach, where everything sits in the window at once.