Planning beats implementation: how to allocate compute budgets across agent teams

Researchers have developed a systematic framework for allocating compute resources in multi-agent teams tackling long-horizon software-engineering tasks, and shown that the stage at which a leader delegates work determines efficiency more than team size does. Experiments on the ProgramBench benchmark demonstrate that a heterogeneous team of 17 agents matches the performance of a strong homogeneous team of 5 agents at 70% lower cost.

The problem: more agents do not guarantee more throughput

Multi-agent systems offer a new scaling axis for inference: instead of a single agent thinking longer or generating more attempts, the task is split among many agents that work in parallel and coordinate. In practice, adding agents creates duplication, coordination failures, and inflated costs. The team — led by Keenan Samway, Daniel Souza, Omar Darwiche Domingues, Mziizm, Yichern Tan, and Nouhad Ziri — framed team design as a compute-allocation problem: starting from a single agent, where should additional compute be routed to reach the reward-cost Pareto frontier?

Methodology: a minimal harness and the Pareto frontier

The researchers built a minimal multi-agent-team harness from scratch and mapped the frontier by varying one factor at a time: the stage at which the leader delegates to workers (planning, implementation, or verification), the models composing the team, and the number of workers. On ProgramBench, teams outperformed a single agent but cost three to seven times more. Proper compute allocation, however, let a heterogeneous 17-agent team approach within 1.4 points of a strong homogeneous 5-agent team at 70% less cost.

Planning: the highest return on investment

The first finding: planning is the most cost-effective use of workers. A leader that delegates only the planning stage to workers, then implements and verifies the code itself, achieves the same reward as a leader that delegates across all three stages — planning, implementation, and verification — at half the cost. When the leader dynamically decides how many workers to spin up, delegating planning alone captures most of the benefit.

Cheap models: good for planning, bad for implementation

The second finding separates roles: cheap models pay off as planners but not as implementers. A strong leader paired with cheap planners recovers 47.5% of the reward gain of an all-strong team for just 13.9% of the additional cost. Cheap implementers, by contrast, degrade reward. The asymmetry is consistent: planners continue to improve reward as the team grows, while implementers and verifiers saturate after four workers.

An agent-budget playbook

The results yield practical guidelines: given task, cost, and latency constraints, invest the bulk of additional budget in planners — preferably cheap ones — and keep implementation and verification in the hands of a single strong model or a small team. The approach turns team design from an educated guess into an optimization problem with a measurable solution.