What if agent teams could reorganize mid-task?

First, the intuition.

A rigid team repeats its habitual pattern even when that pattern fails. A supervisor that recognizes a plateau might introduce a new viewpoint, redistribute responsibilities, or shorten an unhelpful coordination path.

What the researchers actually did

Evolving Agent Teams uses a supervisor to adjust coding-agent roles, interactions, and memory during GPU-kernel optimization. It evaluates tasks on Hopper and Blackwell hardware with executable correctness and speed benchmarks.

What the results suggest

• The system outperformed coding-agent baselines on four primary CUDA optimization tasks under a matched token budget.
• On one MLA paged-decode task, it reached 0.74× a reference kernel at the matched budget versus 0.51× for Claude Code.
• The authors identify cases where fixed teams stalled while restructuring helped.

How convincing is the evidence?

Token-matched solo-agent comparisons, real kernel execution, correctness checks, and a public code artifact support a more concrete argument than anecdotal demos.

What remains uncertain?

Equal tokens do not imply equal wall-clock or tool budget. GPU-kernel optimization is a specialized setting, and several experiments have few repeated runs.

Connections worth exploring

• Adaptive restructuring is a form of meta-control: changing the search procedure when progress stalls.
• Detecting a plateau reliably may be as important as designing the new role.
• Additional coordination can also create new misunderstandings.