AI Teammates: how monday.com runs production AI agents on Amazon Bedrock AI队友：monday.com如何在Amazon Bedrock上运行生产级AI智能体

TL;DR

• monday.com operates a large-scale production AI agent system called Sphera on Amazon Bedrock, integrating agents directly into existing engineering workflows alongside human developers.
• The architecture utilizes a unified event handling system across Slack, GitHub, and monday.com items, leveraging AWS SNS/SQS for reliable message routing and Amazon EKS for agent execution.
• State management is optimized by separating live session state (stored in Amazon ElastiCache for speed) from persistent memory and code workspaces (stored on Amazon EFS to support POSIX-compliant tools).
• The company reports a 50% increase in per-engineer pull request throughput and a rise in AI tool adoption from 50% to 90% among builders within six months.
• The system employs a "confidence-scored merge" strategy and treats agents as teammates with stable identities, enabling them to own end-to-end delivery while remaining accountable to human oversight.

Why It Matters

This case study demonstrates that agentic AI can be successfully integrated into legacy, high-stakes enterprise environments without requiring greenfield development, challenging the notion that AI agents are only viable for isolated experiments. It provides a concrete architectural blueprint for managing AI state, security, and reliability in production, offering valuable insights for organizations looking to scale AI adoption beyond simple chatbots. The significant productivity gains reported validate the business case for investing in robust AI infrastructure and workflow integration.

Technical Details

• Architecture Stack: The system runs on Amazon EKS, using Amazon SNS and SQS for event-driven communication between triggers (Slack, GitHub, monday.com) and agent runners. Model inference is handled via Amazon Bedrock, with AWS Secrets Manager securing per-session credentials.
• State Management Strategy: Live state (task cursors, locks, logs) is stored in Amazon ElastiCache for sub-millisecond access, while persistent session data, code repositories, and memory files are stored on Amazon EFS. This choice ensures compatibility with POSIX-based tools like Git and NPM, which often fail on object storage like S3.
• Unified Agent Identity: Agents are assigned stable identities within a "Teams" interface, allowing them to be tagged, assigned tasks, and code-reviewed just like human employees. This schema integrates seamlessly with existing collaboration tools, ensuring agents operate within established team structures.
• Custom SDK Wrapper: monday.com wraps the Claude Agent SDK with a custom monday-agent-sdk to maintain provider neutrality, reduce cold-start times via pre-warmed containers, and enforce internal evaluation standards and plugin composition rules.
• Reliability Mechanisms: The event pipeline includes built-in retries, dead-letter queues, and durable replay capabilities. Before promoting new builds, the team replays the previous day's events to ensure stability, providing a safety net for autonomous operations.

Industry Insight

Enterprises should prioritize integrating AI agents into existing operational workflows rather than building siloed applications, as this approach yields higher adoption rates and measurable productivity improvements. Investing in robust, low-latency state management solutions like ElastiCache and EFS is critical for maintaining the reliability and POSIX compatibility required by complex coding agents. Finally, treating AI agents as first-class citizens with clear accountability and identity frameworks facilitates smoother human-AI collaboration and easier governance in regulated environments.

TL;DR

• monday.com在十年历史代码库中大规模生产化部署Agentic AI，通过内部系统Sphera实现AI代理与人类工程师的协同工作。
• 采用三层演进策略：从L1辅助编程到L2技能子代理，最终迈向L3多代理自主交付，显著提升工程效率。
• 架构基于Amazon Bedrock及AWS全栈服务（EKS, SNS/SQS, ElastiCache, EFS等），实现高可靠的事件驱动与状态管理。
• 引入“置信度评分合并”机制，逐步缩小与完全自主性的差距，确保在复杂企业环境中代码提交的安全性。
• 内部数据显示，90%的工程团队每月使用AI编码工具，每位工程师的PR吞吐量提升超过50%。

为什么值得看

本文提供了罕见的在企业级遗留系统中成功规模化运行Agentic AI的真实案例，展示了如何平衡创新与稳定性。对于希望将AI代理集成到现有DevOps流程的团队，其架构设计和状态管理策略具有极高的参考价值。

技术解析

• 分层演进架构：monday.com将AI工程分为三个层级：L1作为助手（Pair Programmer），L2构建可复用的技能和子代理（工程师主导），L3实现多代理端到端交付（工程师编排）。目前主要处于L2阶段，显著提升了人均产出。
• 统一事件路由与基础设施：利用Amazon SNS进行事件分发，SQS进行队列管理，EKS运行消费者。这种Pub/Sub加队列的模式提供了自动重试、死信队列、背压处理（应对Bedrock限流）以及持久化重放能力，确保系统在高负载下的稳定性。
• 混合状态存储策略：为优化成本和性能，采用分层存储方案。活跃状态（任务、锁、心跳）存储在低延迟的Amazon ElastiCache中；会话上下文、代码仓库和工作区文件存储在Amazon EFS上，以支持POSIX文件系统兼容性和跨Pod故障转移；敏感信息通过AWS Secrets Manager管理。
• 自定义Harness与SDK封装：虽然底层运行时依赖Claude Agent SDK，但monday.com构建了名为monday-agent-sdk的薄包装层。该层实现了提供商中立性（路由至Bedrock）、冷启动优化（预缓存依赖）以及内部评估标准、插件组合和代码审查逻辑的定制化。
• 身份认同与协作模式：AI代理被赋予类似人类的身份（如角色、经理、绩效分数），融入Slack、GitHub和monday的工作流。代理通过统一的收件箱（@提及、任务分配、PR请求）接收任务，并以“队友”而非“作业队列”的方式参与协作。

行业启示

• 遗留系统的AI现代化路径：证明在拥有数百万用户和十年历史代码库的企业环境中，通过合理的架构抽象（如统一事件总线、分层状态管理）可以安全地引入Agentic AI，无需推倒重来。
• 从自动化到自主化的渐进式采纳：企业应遵循L1到L3的成熟度模型，先在辅助和子代理层面验证价值，再逐步过渡到多代理自主交付，同时保持人类在关键节点（如代码审查、最终合并）的控制权。
• 基础设施即信任基石：在生产环境运行AI代理时，可靠性（重试、重放、故障转移）比模型智能本身更重要。构建稳健的事件驱动架构和状态恢复机制是获得工程团队信任并实现规模化应用的关键。