🗓️ February 28, 2026
📍 Nutanix India, Baner, Pune
Most AI agent demos look impressive in controlled notebooks. But ship them to production, and reality hits hard.Aditya Godbole's upcoming talk at Fifth Elephant Pune 2026 exposes the core mismatch:
- Dev-mode agents thrive on tight context, clean inputs, and single-developer workflows.
- Production agents drown in unbounded telemetry — logs, traces, metrics, events — at massive volume, variety, and velocity.
Raw observability data overwhelms context windows. Noise buries signal. Agents confuse correlation with causation. Investigation paths explode.The fix isn't better prompts. It's structural cognition.Aditya proposes building a schema-driven knowledge graph as the agent's reasoning substrate (not just vector memory for recall). Key ideas:
- Treat metrics as the richest structured source → nodes (labels/entities), edges (metric names/relationships).
- Schema matching + fuzzy entity mapping (via Jaccard similarity) to filter noise and connect telemetry to domain concepts.
- This enables causal reasoning and guardrails that survive real-world chaos.
Vector embeddings help agents remember. Knowledge graphs help them reason.If you're building AI-native SDLC tools or observability-driven agents, this talk (and the related hands-on workshop) offers rare depth across distributed systems, data modeling, similarity algorithms, and agent design. Check out Aditya's submission.
Register for Fifth Elephant Pune 2026.
Production-grade agents aren't about bigger models — they're about smarter memory architecture.
