Cost is the resource use or monetary expenditure associated with operating an agent. It is a design concern alongside predictive performance, interpretability, and alignment. Computational efficiency belongs on the cost side of that distinction: an agent can do its job well while using more resources than necessary. section 1.2.2
One way to reduce cost is to give the agent a process that quickly reaches appropriate resources for a task, rather than encouraging an extensive, expensive search for a plan or solution. Lower cost is a possible benefit of agent design, not a guaranteed one. section 1.1.2
Evaluate cost against what a design change achieves:
- Check whether the change improves the agent’s ability to handle tasks; predictive performance and cost are separate concerns. section 1.2.2
- Account for repeated work. Decompose-Process-Recompose can improve capability, but piece-wise processing usually adds a constant overhead per input fragment. section 8.2.1
- Consider latency separately: processing fragments in parallel can improve latency even when the decomposition adds overhead. section 8.2.1
A large-input agent
An agent splits a large input into fragments, processes each fragment, then recomposes the results. This may let it handle the input more accurately, but every fragment introduces processing overhead. Parallel work may shorten the wait without removing that overhead. section 8.2.1
















