New: Agent Health diagnoses silent failures and offers fixes

Most teams shipping AI agents to production have a monitoring gap they don’t know about yet. Dashboards and logs have their roles, such as giving you a high-level view of agent output performance or error rates. But both of these have limitations; neither tells you that a tool is looping, for example, or that a model fell out of date. Nor will they tell you that a conversation is burning tokens for no reason.

That is the nature of agentic failures. Rarely do you see a dramatic crash; more typically, it’s a slow leak that decays performance and erodes agent trust, such as a stale model version still in production or a prompt that is no longer fit for purpose. And while the calls complete and the outputs look fine, the cost and quality of the agent degrades over time until users stop trusting it.

When there is no signal indicating that something is blatantly wrong, it can be hard for engineering teams to get ahead of this type of failure. Therein lies the challenge in building and maintaining trusted AI.

Monte Carlo’s new Agent Health feature aims to close this gap. Agent Health is an automated insight layer for every registered agent within your ecosystem. It surfaces the single most important issue to fix, backed by evidence, with a concrete recommendation attached — no manual trace-hunting required.

Why most teams fly blind on agent drift

Debugging an agent today usually means someone opening a trace explorer after an issue became severe enough to get noticed, and manually stitching together what happened across dozens or hundreds of spans. While this works when there’s an obvious error, it does nothing for the silent failures that are typical of agents and their workflows.

We ran into this issue ourselves at Monte Carlo many times. You can read about how Monte Carlo’s own Troubleshooting Agent — the agent responsible for automatically investigating data and AI incidents and proposing resolution steps — experienced performance degradations due to an older Anthropic model being used for a specific task, while other tasks had moved on to a newer model. While no clear error was observed, there was a significant efficiency gap and output quality issue that was unnoticed until Agent Health proactively picked it up. You can learn more about this case here.

As in our own example, the information is almost always technically available in an agent’s traces; it’s just that no one is actively looking for it.

This is why Agent Health changes the game.

What you can do now

Agent Health is constantly profiling your agent’s runs so that it can proactively surface issues that are not obvious. It runs continuously against every registered agent and turns raw traces into a prioritized, evidence-backed report which includes the following information:

• A single, ranked top issue — instead of a dashboard full of undifferentiated flags, Agent Health leads with the one problem worth fixing first, with a clear problem statement and the evidence behind it.
• Fused, multi-source evidence — insights are grounded in agent graph and topology, runtime telemetry, error samples, and conversation samples, so findings reflect what the agent actually did, not a guess from a single log line.
• Code-aware grounding — when GitHub is connected, findings are checked against the agent’s actual source, which sharpens root cause and cuts down on false positives from guessing at what the code does.
• Connected to Agent Lineage — the same graph-originated insights also surface in the Lineage tab, so both views stay in sync.
• Fix with AI — every issue comes with a button that packages the full context — the problem, the evidence, the recommendation — into a prompt ready to paste into Claude or Cursor, so getting from insight to fix doesn’t mean starting a debugging session from scratch.

Under the hood, the insights that Agent Health generates are fused from multiple sources for higher confidence. These include agent graph and topology, runtime telemetry (failure rates, latency, retries), real exception types from error samples, multi-turn conversation samples for session-level quality issues, and, when GitHub is connected, actual source code.

Powering up Agent Health with GitHub

Agent Health works without code access, but it gets sharper with it. Connecting GitHub lets Agent Health ground its findings against your agent’s actual source rather than inferring behavior from traces alone — and it’s what unlocks the fastest path from insight to fix, since the recommendation and the “Fix with AI” context can point at the exact file and line involved.

Get started

Agent Health is currently in limited roll-out. If you want a look at what it surfaces on your own agents, get in touch with your Monte Carlo team or schedule a demo.

Our promise: we will show you the product.