Agent Trust Overview

Overview

Agent Trust (Agent Observability in the product) finds what's going wrong with your AI agents and the data that powers them, explains why, and hands you the fix. Every run becomes a trace. Monte Carlo reads those traces continuously, catches the failures real users hit — without you hand-building golden datasets first — and turns each one into a verified, evidence-backed recommendation you approve. Every fix makes the agent better instead of letting it drift.

Trace content — prompts, outputs, and tool calls — stays in your own environment. Monte Carlo receives span metadata and evaluation scores, and your data-sharing settings control access to raw content.

What Agent Trust unlocks

Agent Trust gives you one view across your agents and the data that powers them, so you can ship production-grade AI and catch failures fast.

• Self-improving agents. Monte Carlo analyzes each agent's traces, finds what's failing, slow, or wasteful, verifies it's real and not already fixed, and recommends a concrete fix: a code change, a ticket, or a monitor. You review and trigger it.
• Failures standard observability misses. Tool cycling, retry loops, silent empty output, stale or wasteful model usage, context overflow — failures that return normal status codes and normal latency.
• Agent and data root cause in one place. Every failure is traced across the agent and the data feeding it, with lineage from each output back to its source tables, so you know right away whether it's an agent problem or a data problem.
• Coverage from day one. Register an agent and Monte Carlo sets up monitoring across context, performance, behavior, and output automatically.
• Bayes, Monte Carlo's AI teammate. Ask Bayes to troubleshoot an alert, investigate a finding, or build a monitor — in Monte Carlo, Slack, Teams, or your AI coding assistant through MCP. You decide how much Bayes takes on.

Agent Trust works wherever your agents run: any agent that emits OpenTelemetry — any language, any framework, any model — plus platform-native agents on Snowflake and Databricks. Setup differs by type; everything downstream is the same either way.

How it works

1. Connect your agent — and its definitions, so Bayes can diagnose problems and propose fixes. Setup differs by where it runs, everything after is the same.
2. Bayes takes it from there. Monitoring is set up for you, and Bayes turns recurring problems into diagnosed fixes your team reviews and ships.

1. Connect your agent

Agent definitions let Bayes name the exact cause of a problem and propose a fix. Without them, findings still show what broke, but not why.

OpenTelemetry agents

OpenTelemetry agents send traces to the Agent Observability Data Platform — an OpenTelemetry Collector, ClickHouse for storage, and an LLM worker that runs evaluations — which Monte Carlo queries through the Monte Carlo Agent.

Run it self-hosted in your own cloud with a single Terraform module, or ask your Monte Carlo representative about Monte Carlo hosting and skip to step 3.

Once traces are written, the agent appears on the Agents page on its own.

Platform-native agents

Snowflake and Databricks agents already write trace data on their own platform, so Monte Carlo reads it through your warehouse connection — nothing to deploy.

1. Grant access and enable trace storage — see the setup guide for your platform above.
2. Register the agent. Go to Settings → Agent Observability → Add, set Agent Type to Platform Agent, pick your connection, select the agent, and click Import.

2. Bayes takes it from there

Bayes, Monte Carlo's AI teammate, does the work from here. It sets up monitoring for you and turns recurring problems into fixes your team reviews and ships. You decide how much Bayes takes on

Finds what's breaking your agents and proposes verified fixes

Monitors and evals tell you something broke. With Reinforcement Loop, Bayes tells you why and how to fix it — so improving an agent stops meaning hours spent reading traces one conversation at a time. Every finding moves through the same cycle:

Bayes never changes a running agent — every fix it proposes goes to your team for review before anything lands. Rate any finding and your feedback trains Bayes. Diagnosis and fix proposals need your agent's definitions connected.

Sorts every conversation by what users ask and how it went

Stop sampling conversations at random and by hand. Bayes sorts every conversation as it arrives into clusters by intent (what the user asked for) and by issue (how it went), so you can:

• See what your agent is really used for. Intent clusters show what users actually ask — rarely what the agent was scoped for, and the gap is where to invest next.
• Read the broken conversations first. Issues clusters rank the slow, token-heavy, looping, and low-scoring conversations, so your time goes where the agent is failing.
• Know what to build or grant next. Out-of-scope clusters list what users asked for that the agent couldn't answer — a missing table, document, permission, or tool.

Issues clusters are what Reinforcement Loop works from.

Sets up monitoring for you

After a day of traces, Bayes sets up monitors across the four layers of agent trust — context (the data feeding the agent), performance (slow, costly, or erroring), behavior (tool use and reasoning), and output (the quality of the answer) — limited to what the agent's telemetry supports. Nothing to configure.

For your own rules, like an SLA-driven latency budget, describe the rule to Bayes or to Claude Code, Cursor, or another tool connected through the MCP server with the Agent Toolkit, and it builds the monitor.