From Demo to Production: Why Agentic AI Systems Fail - and How to Fix Them
summary
Claim: agentic failures are a systems problem, not a model one. A platform built for 10M+ users hits four concrete failure patterns: context overload from thousands of instructions bloating 200k-token windows, tool explosion with 200+ tools driving token and latency collapse, agent-to-agent handoffs that lose business rules, and an execution black box with no traces. Fixes used: skill-based progressive disclosure, semantic tool search to load only needed tools, one-agent/multi-skill design, and end-to-end tracing (e.g., LangFuse) for reproducible debugging.
AI-written summary · about
chapters (8)
- 0:00Agentic AI challenges
- 3:55Pilot vs production readiness
- 6:02Context overload in LLMs
- 13:46Tool explosion and search
- 19:47Agent orchestration complexity
- 26:31Debugging execution black box
- 30:45Enterprise AI lessons
- 31:29Q&A
related talks
20mFrom Demos to Production: Why Agentic AI Breaks at Enterprise Scale
20mThe Three Ways AI Agents Fail. Here's How to Fix Yours.
16mBuilding Applications with AI Agents
23mFrom Models to Moments: Why Most AI Products Fail in Week One
53mThe Exponential Value of Agentic AI: Product Lessons from Data to Decisioning
33mFrom Prototype to Production: Making Agentic AI Systems Reliable
33mUnlocking Agentic AI: [...]
34mThe Human-in-the-Loop is Tired