A team of agents can lose time if one planner has to handle every question and change of plan. Coral’s Anemoi research explored giving agents a way to discuss findings, identify bottlenecks and adjust their work together.
More useful work from a smaller planner
The paper reported 52.73% accuracy on GAIA with GPT-4.1-mini as the planner, compared with 43.63% for OWL under the same model settings. GAIA tests an assistant’s ability to solve questions that require reasoning and tools.
For teams choosing AI systems, the result shows why the way agents collaborate deserves attention alongside model choice. Researchers can test different coordination approaches before a team commits more resources to larger models.
Read the paper and study conditions, or explore Coral’s current research.