Coral AI Labs

Research briefs

The results.
The work behind them.

Explore our evaluations of coordinated agents in software and autonomous research.

Coral against the baselines

95 of 124 codebase questions answered perfectly.

SWE-Atlas Codebase QnA results across nine single-agent baselines, with task-level comparisons, costs and methodology.

Explore SWE results ↗

02 · Autoresearch

Coral Tree across nine tasks

Best or joint-best results. Five wins and one tie.

Optimisation experiments comparing Coral Tree with single-agent and four-agent configurations, with scores and visualisations.

Explore Autoresearch results ↗

Shared by direct link. Not listed in our public navigation.