Skip to content

PARTNR

A benchmark of 100,000 multi agent language tasks

Visit official site

What PARTNR is

The benchmark is described as a large scale release from Meta FAIR containing 100,000 natural language tasks for multi agent reasoning and planning research.

What you can do with it

  • Test a multi agent system on the task set
  • Generate tasks with LLMs at scale
  • Evaluate agents alongside real human partners

Application scenarios

  • A researcher tests coordination in complex settings
  • An educator explains multi agent planning
  • A developer tunes an agent for human collaboration

Pros & cons

✓ What we like

  • A simulation loop reduces task errors
  • The material reports 93 percent human and 30 percent LLM task solving

! What to watch out for

  • The stated planner limits cover coordination and error recovery
  • Spatial, temporal and heterogeneous agent constraints are noted

FAQ

How many tasks are included?

The material states 100,000 natural language tasks.

What gap does it report?

The material states humans solve 93 percent of tasks while LLMs solve 30 percent.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review