AWS has published a detailed look at how monday.com runs production AI agents on Amazon Bedrock inside a decade-old enterprise codebase. The case study matters because it focuses on agents working in a live SaaS environment with real users, compliance expectations, on-call responsibility, and code review.
monday.com describes three stages of agent use. Engineers first use AI as an assistant with tools such as Cursor and Claude Code. Teams then create reusable skills and sub-agents for repeated work. The longer-term goal is multi-agent delivery, where agents pick up tasks, communicate in Slack and monday.com, and ship code alongside humans.
The company’s internal system, Sphera, treats agents as teammates with stable identities, managers, scopes, and performance scores. An agent named Atlas, for example, has the role of software engineer and can be assigned tickets from the same backlog humans use.
The point is not that agents are fully autonomous everywhere. The architecture shows the management layer needed before agents can touch production: identity, accountability, review, confidence scoring, and clear ways to deactivate or limit them when they fail.