Monte Carlo ships a new observability tool for agents, because LLMs are hard to read

Monte Carlo released a tool that watches agents — the kinds that use tools, call APIs, and chain together several model calls — and records what they do so engineers can actually debug them. For a while, observability was built for APIs and microservices: latency, errors, traces. Agents moved the goalposts. A single request can spawn a dozen calls to different models, each with different prompts and different tool calls. The old tools don't capture that.

The new tool sits between your agent and the models it calls. It logs every request and response, ties them together into the agent's execution graph, and surfaces the data in a dashboard. What you get is the ability to replay an agent run and see exactly which model call produced which result — not just the final answer, but the path there. That's the thing that makes agent debugging possible instead of guesswork.

Monte Carlo has been in the observability game for a while, mostly around data quality and pipelines. This moves them into the agent layer, where the same problems show up in a new shape. If your agents are failing silently or making the wrong tool calls, this gives you the traces to find it.

Why this matters for us: agents are going to be running the tools la gente uses — scheduling, payments, searches, the whole stack — and when they break at 2am, you need to know why without writing a new debugging tool.