monday.com has implemented AI agents on Amazon Bedrock to enhance productivity, with internal data showing a significant increase in developer efficiency. The company reports that nine in ten Builders use AI coding tools every month, up from roughly half a year ago. Per-engineer PR throughput has also risen by more than half, according to monday’s internal production data. This post outlines the architecture behind these improvements, the challenges of integrating AI into a decade-old code base, and the confidence-scored merge strategy that is closing the gap to full autonomy. The system is designed to handle real-world demands, including on-call responsibilities, customer interactions, and compliance, which makes it distinct from greenfield demos. The AI agents operate within the company's existing infrastructure, ensuring seamless integration with millions of users and hundreds of microservices. This approach allows for real-time collaboration and accountability, treating agents as teammates within the organization. Source: awsml
The implementation of AI agents at monday.com involves three levels of engineering, starting with AI as a pair programmer for fast reflexive tasks and heavy lifting. Engineers use tools like Cursor and Claude Code, with adoption nearly doubling year over year. At the second level, teams build reusable agents for repeated work, significantly boosting per-developer PR throughput. The third level involves fully agentic agents that take ownership of delivery while engineers orchestrate tasks, integrating with platforms like Slack and monday. These agents are part of monday’s internal system, Sphera, which manages a mix of human and agent teammates with profiles, managers, and performance scores. The agents, like Atlas, operate with a stable identity across Slack, GitHub, and monday, allowing for tagging, assignment, and code reviews as if they were human colleagues. Source: awsml
The architecture of the AI agent system at monday.com is built on seven AWS services, including Amazon SNS, SQS, EKS, RDS, ElastiCache, EFS, and S3. These services support the event path from external triggers to the execution of agent tasks, ensuring reliability and scalability. The system uses a combination of pub/sub and queue mechanisms to manage retries, dead-letter queues, back-pressure, and durable replay of events. The monday-agent-sdk serves as a thin wrapper around the Claude Agent SDK, providing provider neutrality, cold-start cost reduction, and a custom harness for evaluating agents, managing plugins, and interacting with external platforms. The state and memory of agents are managed through ElastiCache and EFS, with durable records stored in S3 for audit purposes. Amazon Bedrock is central to the system, enabling the execution of hundreds of agents while maintaining cost, safety, and capacity control. Source: awsml
