Session abstract
What you’ll learn
[Disclaimer: The opinions, insights, and perspectives shared in this session are solely my own and do not reflect those of my current or past employers, nor do they convey any confidential information by any means.] Your agent nailed the demo, then fell apart in the real world, and prompt-tuning didn't fix it. MIT's 2025 NANDA study found 95% of enterprise GenAI projects deliver no measurable return, and the cause is less often the model: it's the system around it. Those failures cluster in three layers, and this talk walks through all three, grounded in one running example: an agent modernizing a legacy enterprise system. Layer 1: Tools (MCP and tool-schema design) An agent is only as reliable as the tools it calls. We cover why vague tool definitions cause silent, hard-to-debug failures, and the design choices (typed inputs, clear error contracts, idempotency) that make a tool call dependable. Layer 2: Memory (RAG vs. KG) An agent is only as good as what it remembers. Retrieval-Augmented Generation (RAG) pulls snippets but misses how they relate, so an agent modernizing a legacy monolith drops a rule buried in its dependencies that a Knowledge Graph (KG) would catch. We cover when to use each, or both. Layer 3: Coordination (orchestration and agent-washing) More agents is not more reliability. We separate real orchestration from "agent-washing" and cover what stops one failure from cascading: bounded autonomy, failure isolation, and human sign-off on irreversible steps - the "risk controls" Gartner ties to 40% of agentic cancellations by 2027. Three-quarters of teams are adopting agents; only a small minority reach production (Forrester, 2026). This talk is about joining them. You leave with: 1. The Failure Map: the three layers where agents actually break 2. Go-To Schema: design choices that stop silent failures 3. RAG vs. KG: when to use each, and when to combine them 4. No-Cascade Coordination: isolate failures, spot agent-washing