GenAI Grows Up: Building Production-Ready Agents on the JVM

Rod Johnson’s talk provides a compelling and pragmatic perspective on the challenges of deploying GenAI in enterprise environments. His credibility as the creator of Spring Framework lends weight to his arguments, and he effectively articulates the gap between the hype and the reality of production AI systems. The talk is well-structured, moving from the promise of GenAI to the specific reasons for enterprise failures, and then to a proposed solution. Johnson’s emphasis on non-determinism as a fundamental challenge is astute, and his suggestion to break down complex tasks into smaller, orchestrated steps is a practical and widely accepted approach. He also correctly identifies integration as a major hurdle, noting that most business logic resides on the JVM, not in Python, which is a valid point often overlooked in AI discussions. The introduction of his framework, Embabel, is a natural culmination of his argument, though it inevitably carries a promotional element. However, the talk lacks empirical evidence or case studies to substantiate the claimed failure rates or the effectiveness of his proposed solutions. While Johnson references the MIT study on 95% failure, he dismisses it without deeper analysis, and no other studies are cited. The discussion of prompt engineering as ‘alchemy’ is vivid but somewhat dismissive, and he does not delve into emerging techniques like prompt optimization or few-shot learning that might improve reliability. The focus on the JVM is justified, but the talk could have benefited from a more balanced discussion of alternative platforms or hybrid approaches. Overall, the talk is insightful and valuable for practitioners, but it is more of an expert opinion than a rigorous scientific analysis. The adéquation between title and content is strong, and the talk does not suffer from any significant digressions. The presence of a promotional segment for Embabel is transparent and does not detract from the core message.

The title accurately reflects the content: a discussion on moving GenAI from experimentation to production, with a focus on JVM-based agents.

8/10

Rod Johnson, creator of Spring Framework, provides a well-argued, experience-based perspective on enterprise GenAI. He cites specific examples (Air Canada) and references industry trends, but the talk is primarily opinion and advocacy for JVM-based agents, lacking empirical data or peer-reviewed sources.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction by Sam and Rod Johnson's opening remarks.
• 01:53Rod Johnson introduces himself and his background.
• 02:24Discussion on how tools add power to GenAI.
• 03:36Personal assistant use case and its success.
• 06:00Explanation of tool calling and MCP standard.
• 08:00Emergence of open models and their enterprise relevance.
• 10:00Comparison between personal assistance and business process automation.
• 12:00Reasons for enterprise GenAI failures: non-determinism, hallucinations, integration.
• 15:00Air Canada chatbot hallucination example.
• 18:00Discussion on prompt engineering as 'alchemy'.
• 21:00Importance of orchestration and breaking down tasks.
• 24:00Integration challenges and the role of JVM.
• 27:00Top-down vs bottom-up approaches in enterprise AI.
• 30:00Introduction of Embabel framework and its principles.
• 33:00Emphasis on testability and maintainability in AI systems.
• 36:00Conclusion and call to action for building production-ready agents.

• Rod Johnson's GitHub — Repository for Embabel and other projects.
• Rod Johnson's LinkedIn — Professional profile.
• Rod Johnson's blog — Personal blog with articles on software and AI.
• GOTO Copenhagen 2025 session page — Session details for this talk.

• GOTO Conferences on Bluesky — Social media presence of the conference.
• GOTO Conferences LinkedIn — Company page for GOTO.
• GOTO YouTube channel — Channel where the talk is published.

• gotocph.com
• gotopia.tech
• gotopia.tech