Budget first, agents second: structuring multi-agent systems under a hard cost ceiling
What would your agent team look like if the budget came first?
Opening
Most multi-agent systems are designed from the ambition side: choose the capable models, draw the topology, and only then discover what the month costs. The paper starts from the other end. As these systems grow in complexity, cost becomes an important part of deploying them for real, and the authors point out that almost nobody works out how to structure a multi-agent system inside an explicit budget. For a team that quotes a customer a fixed monthly spend before choosing models, that gap is the whole design problem.
What the paper does
BAMAS builds a system in three moves. First, it chooses an optimal set of LLMs by formulating and solving an integer linear program that trades performance against cost. Second, it decides how those models collaborate by using a reinforcement learning method to select the interaction topology. Third, the system is instantiated and executed from the chosen agents and topology. The authors evaluate BAMAS on three representative tasks and compare it with state-of-the-art agent construction methods. The reported result is comparable performance while the cost falls by as much as 86%.
Findings
Budget-shaped design is an open problem — The starting observation is that cost gets treated as a reporting problem rather than a design input, which is why published systems are usually structured with no budget in the room at all.
existing work rarely addresses how to structure multi-agent systems under explicit budget constraints
Model choice becomes a solved calculation — Instead of ranking favorite models by hand, selection is posed as a discrete optimization: a solver returns the set that balances performance against cost under the ceiling it is given.
an Integer Linear Programming problem that balances performance and cost
The same performance for up to 86% less money — That is the trade the paper puts on the table: the budget-aware construction matches state-of-the-art quality on the tasks tested while cutting spend by as much as 86%.
BAMAS achieves comparable performance while reducing cost by up to 86%
Caveat
The authors state their scope plainly: BAMAS is evaluated on three representative tasks and compared with state-of-the-art agent construction methods. The 86% figure and the comparable-performance claim hold on those three tasks. Outside them the paper reports nothing, so a team adopting the method carries the evaluation burden for its own task mix.
What this means for 5ac
The budget at 5ac is known before the pipeline runs: a customer sets a monthly ceiling, then cards get routed. Treat that ceiling the way BAMAS treats it, as an input to model selection rather than a report read afterward. Pick the model per card type as the cheapest one that still clears the quality gate, then let the remaining budget decide which profiles talk to which. Drafting cards go to the cheap lane, strategic pillar pages to the heavy model. When the month tightens, cut the topology first: drop a reviewer step or a child card before touching quality. Cost per task belongs on the card as a design parameter, not on an invoice.
Source
- BAMAS: Structuring Budget-Aware Multi-Agent Systems — https://arxiv.org/abs/2511.21572