Monte Carlo

Breaking: Monte Carlo just launched Bayes!!!! An AI teammate that watches your AI agents and the data feeding them, finds what broke, and proposes the fix. The detail that got my attention: Bayes will draft the right fix for engineers on its own and file it as a Jira ticket or PR - all you have to do is click "approve." That's as close to self-healing as it gets Why this lands for me. This fall I have been on the floor at Dreamforce, UiPath FUSION, Equinix Horizon, IBM Think Mumbai and Gartner IT Symposium. Almost every conversation with a CIO or CDO hits the same wall. The agent works as a pilot. Then it goes to production, a pipeline lands late at 2 AM, and the agent keeps answering confidently from stale data. Nobody notices until a customer does. And the on-call team? Half of them muted the alert channel months ago. Monte Carlo’s launch example is painfully familiar: 12 alerts between 2:50 AM and 4:29 AM. A late Stripe load, a timed out dbt rebuild, missing rows, a slow support agent. At 7:42 AM, Bayes posts 1 message: - Support agent is quoting yesterday’s refund amounts. - Cause: the orders rebuild timed out after the Stripe load arrived 50 min late. - Fix: re-run orders and 2 downstream models, raise the timeout to 45 min. - 4 alerts grouped into 1 issue. 1 agent, 3 tables. That is root cause traced across the agent AND the warehouse in one place. In most enterprises I talk to, those live in 2 tools owned by 2 teams. What is under the hood: 1. Learn: reads your AI and data stack, proposes a monitoring plan, deploys it when you approve 2. Detect: triage, is it real and how urgent 3. Fix: evidence-backed root cause and a suggested remediation 4. Measure: verifies the fix actually landed, then loops again Add monitor tuning so coverage does not rot, and PII discovery for compliance. The part enterprise leaders should care about most is the autonomy ladder: Manual, Assisted, Supervised, Autonomous. Bayes earns scope domain by domain. Start at “it proposes, I approve.” Expand when it proves itself. Early numbers Monte Carlo shared: 1 global bank: 200%+ more real incidents caught, 87% fewer false positives 1 global retailer: 75% of monitoring deployed by the agent, full coverage in 30 days 1 beverage distributor: plan to production in 24 hrs It works in Slack, in the Monte Carlo platform, and via MCP inside Claude Code and Cursor. My take: last year everyone was racing to build agents. This year the hard question is whether you can prove why your agent said what it said. Fixing the agent without fixing the context it relies on it only moves the problem. For leaders running agents in production: what level of autonomy would you give a teammate like this on day 1? See it here: montecarlo.ai/bayes #montecarlo #aiagents #dataobservability #agentobservability #dataengineering #enterpriseai #aitrust #theravitshow

In this month’s Human in the Loop, we’re introducing our newest teammate, Bayes! Bayes is the world’s first AI teammate to deliver self-improving agents and the data powering them. It catches the failures your real users hit, fixes them, and gets better at your environment with every fix, so your team can trust its agents at the speed they ship. We also get into: 🔑 Why so many teams are taking the keys back from their agents 🚨 Another rough month for agents (one of them sent police a fake murder tip, seriously) 🎙️ Barr and Lior on DM Radio, talking about keeping agents trustworthy 💡 The meme of the month Give it a read and subscribe to stay on top of how teams are building agents they can trust. 👇 👇 👇

Excited to see our Bayes launch covered in SiliconANGLE & theCUBE! Self-improving agents let you scale trusted agents in production. And, if you don't follow our co-founder and CTO Lior Gavish, you might want to!

Trust moves at human speed. Agents don’t. That gap is why we built Bayes, Monte Carlo’s AI teammate. It catches the failures real users hit, traces them through your agents and the data that powers them, fixes the root cause, and confirms the fix. Resolutions, not more alerts. You decide how much Bayes takes on. Barr has the full story below.

Big news coming Wednesday .... anyone building agents in production will want to see what we're unveiling! Stay tuned!

Our CEO Barr Moses took the TechCrunch Disrupt stage with Rebecca Bellan, Cathy Gao, Aaron Jacobson to talk about the future of SaaS, AI-era moats and pricing. Give it a listen to hear more about how Monte Carlo transformed to be AI-first!

Curious how we help enterprises trust their agents at scale AND keep an eye on agentic spend? Take a peak during a live demo we're hosting later this week!

More agents in production. More ways for them to fail quietly. That's why we're relaunching our newsletter as Human in the Loop. Every month you'll get the agent news that matters, lessons from teams running agents in production, and what we're shipping to help you trust yours. Subscribe here 👇

Agents spend money on every run, and most teams find out how much when the bill arrives. Tying that spend back to a step, a tool call, or a retry loop is manual work: pull traces, export token counts, guess which change made things worse. Sign up for this live demo on October 7 to see how Monte Carlo helps you learn more about what your agents cost, where that spend is wasted, and what to do about it!

How do you build a roadmap for a category with no playbook? Omer Spillinger on Monte Carlo's agent observability team shared our approach with Built In: start with failure modes you can point at in production. An agent that loops fifteen times and burns a million tokens while reporting success. An agent that answers a revenue question without ever querying the database. Those earn roadmap space, because they're what erodes trust in agents at scale. Every bet is an experiment with an owner, a ship date, and an honest readout. And we run the product against our own agents, so we live with our decisions before customers do. Full piece linked below! #AgentObservability #ProductManagement