Rifft
Debugger for multi-agent AI systems
I've been building multi-agent AI systems for a while and kept hitting the same frustrating problem — something breaks in production and you have no idea why.
Standard logging shows you the final output. Maybe a stack trace. But when you have multiple agents handing off to each other, that's not enough. You need to know which agent made the bad decision, what it was thinking, and what it passed downstream.
I looked for tools that could help and found most observability products are built for single model monitoring — latency, token counts, prompt/response pairs. Not for the failure patterns that emerge when agents interact with each other.
So I built Rifft. It traces every agent decision and handoff using OpenTelemetry, classifies failures automatically using the MAST taxonomy (a UC Berkeley framework for categorising agent failure modes), and lets you replay a broken run step by step without restarting the whole pipeline.
Launched it this week. Free tier available. Still very early but would love feedback from anyone building multi-agent systems — especially if you've hit this problem and found a way to solve it.
2 Comments
Standard logging usually fails in multi-agent workflows because it captures individual outputs without showing the logical chain of thought that led to a specific error. A useful insight is that agent hallucinations often stem from a misalignment in the system prompt of the second or third agent in the chain rather than the initial query. Have you thought about adding a feature to visualize the agent handoff graph in real time to spot bottlenecks or loops?