Runloop now integrates with @wandb Weave for orchestrated agent benchmarks with full traceability. Runloop runs thousands of agent tasks in parallel. @weave_wb turns the traces into something you can inspect and compare. Joint report: wandb.ai/wandb_fc/genai…

@wandb @weave_wb Agent benchmarking at scale has two problems: 1. Most benchmarks don't run in parallel, so evaluation takes days 2. The output is a pile of logs nobody can read Runloop solves the first. Weave solves the second.

What the integration looks like in practice: Runloop orchestrates concurrent devboxes, materializes deterministic inputs, isolates the scoring harness, exports structured traces. Weave ingests those traces and provides tool call trees, error clusters, version comparisons, model leaderboards.

@wandb @weave_wb The demo in the joint report: Terminal-Bench 2, OpenCode as the agent harness, Gemini 3 Pro vs Claude Sonnet 4.6, 100 concurrent devboxes, full trace export to Weave, side-by-side comparison in one view.

@wandb @weave_wb Runloop handles the execution layer. Weave handles the observability and analysis layer. Together they turn one-off benchmark scripts into a continuous evaluation workflow. Full report: wandb.ai/wandb_fc/genai…