Monte Carlo, which sells software that watches data pipelines and AI agents for errors, launched Bayes on October 28. The company calls it an AI teammate: it reads how a customer's AI agents behave in production, traces failures back to the data they use, and proposes fixes, according to Monte Carlo's announcement.
Monte Carlo customers can use Bayes now in the company's platform, in Slack and Microsoft Teams, and through MCP, the company said. MCP connects it to AI coding assistants such as Claude Code and Cursor, according to the Bayes product page. Monte Carlo has not published a price.
The pitch is about staffing. Monte Carlo said that across its customers, a typical AI engineer can properly watch only one to three agents, and Barr Moses, its co-founder and CEO, said "there aren't enough engineers to keep those agents trustworthy."
How Bayes works
Bayes reads production traces, meaning the record of an agent's tool calls, model calls, timing, errors, retries, and evaluation scores, according to Monte Carlo's launch post. Bayes then looks for failures the company says standard monitoring misses: empty or partial answers that still return a normal status code, agents stuck retrying a call, oversized models spending tokens, slow steps, and answers at odds with the agent's own data. Monte Carlo said every finding starts with deterministic detectors, and language models only interpret the evidence those detectors gather.
Before it reports an issue, Bayes checks it against runtime evidence and, where connected, the customer's source code, and looks at recently merged pull requests to see whether someone already fixed it. Fixes arrive as draft pull requests, Jira or Linear tickets, or new Monte Carlo monitors, and Bayes reports back when a fix has worked.
On the data side, Bayes picks the tables and pipelines the agents depend on, sets up monitors within a budget the customer sets, and folds alerts that share one upstream cause into a single incident. Customers choose how much it does alone, across four levels: helping when asked, doing an assigned job for review, proposing fixes for approval, or running in the background and raising only exceptions.
What is new for Monte Carlo
Monte Carlo announced two AI agents in April 2025, according to its announcement at the time. Both worked on data quality and left the fix to people. Bayes covers the agents too, writes the proposed change, and checks whether it worked.
Product | Announced | Watches | Delivers |
|---|---|---|---|
Monitoring Agent | April 2025 | Data | Recommended monitoring rules and thresholds |
Troubleshooting Agent | April 2025 | Data | Root cause and recommended next steps |
Bayes | October 2026 | Agents and their data | Draft pull requests, tickets, or monitors, then a check that the fix worked |
Source: Monte Carlo's announcements of April 17, 2025, and October 28, 2026, and its Bayes launch post.
Monte Carlo said customers using Bayes have caught 200% more incidents than their in-house systems did, with over 80% fewer false positive alerts. The company's launch post ties those figures to one financial institution, where one engineer used Bayes to put monitoring on 88% of the company's agents. The numbers are Monte Carlo's own, and SaasBuzz has not tested Bayes.
What Monte Carlo has not said
The price is the main gap. Monte Carlo's pricing page sells credits across four tiers, from Start to Business Critical, with the cost per credit available on request, and does not list Bayes in any tier. The launch materials also do not name the language models Bayes runs on, or say how many customers used it before launch.
