Why Agents Don’t Ship: Closing the Production Gap for Real Business Impact
summary
Most agent demos never reach production because teams skip engineering the pipeline parts that actually make them reliable: the talk shows that building tracing of intermediate steps and production telemetry lets teams debug multi-tool agents and find which retriever or tool fails, that LLM-as-judge evaluation gives cheap, automated multi‑dimension scoring (relevance, grounding, safety) and that scaling needs an “agent mesh” — domain agents as governed interfaces — not ad hoc replicated data silos.
AI-written summary · about
chapters (10)
- 0:00Agent production challenges
- 2:06Why agents rarely reach production
- 4:58Customer progression with agents
- 7:07ML Ops vs GenAI Ops
- 9:13AI tracing in GenAI apps
- 12:24Evaluating outputs with LLM judges
- 15:10Multi-dimensional agent evaluation
- 17:18Scaling with agent mesh
- 22:34Recap and scaling transition
- 23:15Q&A on agents and mesh
related talks
34mThe Hard Truth About AI Agents in Production
40mHow To Tame Your Agents?
28mAI Agents as Architectural Strategists | Aligning System Design with Business Goals
39mFrom Demo to Production: Why Agentic AI Systems Fail - and How to Fix Them
02h 04mShip Real Agents: Hands-On Evals for Agentic Applications
30mAgents Building Agents- 02h 04mShip Real Agents: Hands-On Evals for Agentic Applications
07h 37mAI Agents Conference 2026 - Day 3