Agentic Evaluation Authoring#

Arena v0.4 treats coding agents as a primary entry point for authoring evaluation environments. These two examples show agents using Arena’s placement relations and composite tasks to build workbench cleanup and kitchen pick-and-place evaluations.

Workbench
Ten reset layouts of the Arena workbench cleanup environment

An agent authors a workbench cleanup evaluation environment.

Workbench Environment
Kitchen
Ten reset layouts of the Arena agent-authored peninsula kitchen environment

An agent authors a kitchen evaluation environment with a composite pick-and-place task.

Kitchen Environment