monday.com has deployed AI agents, called AI Teammates, in production on Amazon Bedrock, achieving significant productivity gains. Nine in ten Builders use AI coding tools monthly, and per-engineer PR throughput has increased by over 50%. The company's internal agent system, Sphera, treats agents as teammates with profiles, managers, and performance scores. The central agent, Atlas, picks up tickets, writes PRs, and ships features from the same backlog as human engineers.
Architecture
The system uses three inboxes—Slack @mentions, monday item assignments, and GitHub PR review requests—all feeding into the same agent session. Events flow through Amazon SNS to per-team SQS queues, consumed by the monday Builders CoWORK on EKS. The monday-agent-sdk wraps the Claude Agent SDK for provider neutrality, cold-start optimization, and custom harness logic. State management uses Amazon ElastiCache for live state, Amazon EFS for sessions and memory (including cross-session memory files), and Amazon S3 for durable records. Amazon Bedrock provides model calls with cost tracking, audit trails, cross-region failover, and VPC isolation via PrivateLink. Compute runs on Amazon EKS with one pod per active session, auto-scaled by KEDA.
Retrofits for a Decade-Old Codebase
Five key retrofits enabled agent effectiveness: (1) a confidence-scored merge play that gates autonomous merges based on test results and review confidence; (2) per-session secret management via AWS Secrets Manager; (3) a shared EFS filesystem for workspace and memory, allowing session resumption across pods; (4) durable replay of events against patched builds before promotion; and (5) integration with existing Slack, monday, and GitHub workflows without disrupting human processes.
monday.com emphasizes that running agents in an enterprise SaaS requires careful architecture, not just greenfield demos. The system handles millions of users and hundreds of microservices, with agents accountable for real work alongside human teammates.