What happened
Weights and Biases launched W&B Weave Agents on July 25, extending its Weave observability platform to multi-step agent workflows. The product traces tool calls, model hops, latency and token cost, and lets teams score agent runs against curated golden tasks before promoting a version.
Early customers include DoorDash, Canva and Notion. Integrations cover LangGraph, CrewAI and custom MCP servers, with export to existing BI and incident tools.
Why it matters
Agent demos fail quietly in production because teams lack the same eval discipline they built for single-shot LLM prompts. A dedicated tracing and regression layer is becoming infrastructure, not a nice-to-have.
Shipping eval gates before rollout is the same pattern that stabilized traditional software releases, now applied to autonomous workflows.
MintedBrain take
If you run agents beyond a pilot, treat eval datasets and trace review as part of the release process. Weave Agents is one option; the important part is blocking promotion when golden tasks regress, regardless of vendor.
Discussion
Sign in to comment. Your account must be at least 1 day old.