Agents in Production: Lessons shared at AI in Production 2025
Agents in Production: Lessons from the Field
We're excited to share insights from Hannes' recent presentation on deploying AI agents in production environments. As a Principal Machine Learning Engineer with a decade of experience across fintech, healthcare, HR, and retail, Hannes brings real-world perspective to one of the hottest topics in AI today.
The Reality Check: Agents vs. "Process Daemons"
One of the key takeaways? The term "agent" might be doing more harm than good. Hannes argues we should call them "Process Daemons" instead - a name that better sets proper expectations.
Infrastructure That Actually Works
Hannes shared hard-won lessons about agent infrastructure:
Framework Reality: Open source frameworks like LangChain and CrewAI are great for prototyping but bring too many dependencies for production. The recommendation? Implement your own core agent loop.
Tool Integration: Rather than manually defining every tool, they use Go's reflection to dynamically generate JSON schemas from existing APIs - letting existing access controls handle security.
Memory vs. Storage: A crucial distinction that many miss. Use memory as a tool rather than relying on provider memory to avoid vendor lock-in.
The Full Picture
The presentation walks through building a complete agent system with observability, guardrails, memory, and retrieval - showing how what starts as "100 lines of code" quickly becomes a sophisticated system requiring careful architecture decisions.
Key Recommendations
- Prioritize observability and guardrails from day one
- Let applications drive infrastructure decisions, not the other way around
- Focus on responsible deployment with feedback mechanisms and monitoring
- Consider task planning with reasoning models for better performance
- Use reinforcement learning to continuously improve agent-specific models
The Bottom Line
As Hannes puts it, most AI products today are "basically a chat bot on top of old software." The companies winning are those taking a systems-level approach - rearchitecting how data, context, and control flow together rather than just layering AI on top of legacy systems.
For teams serious about putting agents into production, this presentation offers a practical roadmap based on real experience building and deploying these systems at scale.
Want to dive deeper? The full presentation covers implementation details, demos, and specific technical recommendations for each component of a production agent system.
