DeepMind’s 180-agent study says teams win on split tasks, but lose on sequential work
brucemacv · x · 2026-07-23
A thread distilled a Google DeepMind result into concrete agent-design advice:
- The researchers tested 180 agent team setups under the same budget.
- For work that can be split into independent chunks — research, audits, broad scans — teams performed much better, with one cited result showing 80.9% better than a single agent.
- For step-by-step tasks where each action depends on the previous one, every team variant lost to a single agent.
- The takeaway: don’t add agents just to add agents. Whether a workflow should use one agent or many depends on the shape of the work, and any merge step needs a clear owner to avoid error amplification.
More from coding & agent
- GLM 5.2 Reportedly Handles Multi-Node H100 Clusters, Emerging as Top Open-Weights Autoresearch Model — _ScottCondron · 2026-07-23
- For enterprise AI agents, the first question is ownership and governance, not Copilot Studio vs. Foundry — WirelessLife · 2026-07-23
- Tweaking prompts and state to make a model act from vision alone — mervenoyann · 2026-07-23
- AI-written PowerShell script and Codex fix a laptop camera issue in 5 minutes — flowersslop · 2026-07-23
- Agent planning in a generated world may need external memory, not just frames — RecognitionBorn9180 · 2026-07-23
- Agent-native design canvases map out Figma, Paper, and OpenPencil roles — zakelfassi · 2026-07-23