This podcast transcript provides valuable insights into the production deployment challenges and solutions for AI agent systems, featuring perspectives from Joe Mora (CEO of Crew AI) and Vicram Chattery (CEO of Galileo). The discussion reveals real-world experiences with enterprise LLMOps implementations and the emerging technical stack required for multi-agent systems. **Company Context and Use Cases** Crew AI has established itself as a leading orchestration platform for multi-agent systems, with their open-source framework gaining significant traction in the developer community. The company is experiencing rapid enterprise adoption, with customers deploying multiple agent crews simultaneously. A particularly striking example comes from a top-10 US bank that has already built and deployed multi-agent systems across their operations, demonstrating the accelerated adoption timeline in highly regulated industries. Galileo provides observability, evaluation, and reliability tools specifically designed for LLM applications and agent systems. Their platform focuses on custom evaluation frameworks and automated insights generation, moving beyond traditional "magic metrics" to business-specific reliability measures. **Technical Architecture and Stack Evolution** The conversation reveals the emergence of a distinct "AI agent stack" that extends beyond traditional LLM infrastructure. This stack includes several critical layers that organizations must consider when moving from prototype to production: At the foundation level, traditional data management platforms like Databricks, Snowflake, and BigQuery remain relevant. However, the middleware layer requires new components specifically designed for agent orchestration, including memory management systems, authentication frameworks, and inter-agent communication protocols. The authentication layer presents particular challenges for production agent systems. Unlike traditional applications, agents often require dynamic scope changes during execution based on the specific tasks being performed. This creates complex security scenarios where agents may need access to different resources at different times during their workflow execution. **Production Deployment Challenges** Organizations face significant challenges when scaling from prototype agents to production multi-agent systems. The non-deterministic nature of agent behavior creates new categories of failure modes that traditional software engineering practices don't adequately address. Testing becomes particularly complex because unlike traditional software with predictable input-output relationships, agent systems can produce varying outputs for identical inputs. The conversation highlights how enterprises are deploying multiple agent use cases simultaneously - with some advanced customers launching five new use cases per month. This rapid deployment pace creates operational complexity around monitoring, governance, and system maintenance that requires new tooling and processes. **Evaluation and Observability Framework** Traditional LLM evaluation approaches prove insufficient for complex agent systems. The discussion emphasizes moving beyond generic metrics toward custom evaluation frameworks tailored to specific business use cases and regulatory requirements. This is particularly critical in regulated industries like financial services, where audit trails and compliance monitoring are mandatory. Galileo's approach involves creating high-accuracy custom evaluators that can be integrated into CI/CD pipelines. Their Luna family of models specifically targets agentic evaluation scenarios, providing automated failure mode detection and optimization recommendations. This represents a shift from reactive monitoring to proactive system improvement. **Governance and Regulatory Compliance** Financial services organizations are moving surprisingly quickly to adopt agent systems despite heavy regulatory constraints. The key to successful implementation lies in building governance frameworks directly into the agent architecture rather than treating compliance as an afterthought. Risk management teams require visibility into agent decision-making processes and the ability to enforce policies programmatically. This necessitates comprehensive audit logging, decision traceability, and the ability to demonstrate compliance with existing organizational policies through automated evaluation systems. **Interoperability and Protocol Standards** The industry faces critical decisions around standardization of agent communication protocols. While Microsoft's MCP (Model Context Protocol) provides some standardization for tool calling, multi-agent communication protocols remain fragmented. The conversation suggests that successful standardization will require industry-wide collaboration rather than proprietary approaches by individual cloud providers. There's particular concern about cloud providers using protocol ownership as a market control mechanism, which could fragment the ecosystem and limit interoperability. The speakers advocate for open-source, collaborative approaches to protocol development similar to successful financial services interoperability standards. **Development Workflow Evolution** The traditional software development lifecycle requires significant adaptation for agent systems. The non-deterministic nature of agent behavior means that standard testing approaches often fall short. Organizations need to implement continuous evaluation frameworks that can detect behavioral changes and performance degradation across agent fleets. LLM-as-a-judge evaluation becomes crucial for measuring agent performance, but must be complemented by custom metrics that align with specific business objectives. The development workflow must accommodate rapid iteration while maintaining production stability through robust evaluation pipelines. **Democratization Through No-Code Platforms** Crew AI's strategy includes democratizing agent development through no-code interfaces, enabling non-technical users to create and deploy agent systems. This approach aims to scale agent adoption beyond technical teams while maintaining the robustness required for production environments. The challenge lies in making agent development accessible while preserving the control and observability required for enterprise deployment. This requires sophisticated abstraction layers that hide complexity while providing necessary governance capabilities. **Scale and Performance Considerations** As organizations deploy hundreds or thousands of agents, new operational challenges emerge around resource provisioning, load balancing, and system orchestration. The infrastructure requirements for agent systems differ significantly from traditional applications due to their dynamic resource consumption patterns and complex inter-agent dependencies. Performance monitoring becomes more complex as success metrics must account for multi-step workflows, tool invocations, and handoffs between different agents. Traditional application performance monitoring tools are insufficient for these use cases. **Future Outlook and Market Transformation** The discussion suggests that agent systems will fundamentally reshape how software applications are built and operated. Organizations are beginning to question traditional workflow assumptions as agents automate complex multi-step processes that previously required human intervention. The pace of adoption is accelerating beyond typical enterprise technology adoption curves, driven by clear ROI demonstrations and competitive pressure. This creates opportunities for early movers while requiring rapid adaptation of supporting infrastructure and operational practices. The conversation concludes with recognition that the agent ecosystem is still in early stages but moving toward enterprise-grade maturity rapidly. Success in this environment requires balancing aggressive innovation with robust operational practices, particularly around evaluation, observability, and governance frameworks that can scale with organizational adoption.