Session

Observability's Agentic Moment: Operating AI in Production with Genie Code

Overview

Experience	In Person
Track	Artificial Intelligence & Agents
Industry	Enterprise Technology, Healthcare & Life Sciences, Financial Services
Technologies	Genie
Skill Level	Intermediate

Shipping AI products is easy. Trusting and operationalizing them at production scale is the hard part. Operational observability is now having its agentic moment. Join us to see how Databricks is reinventing operational observability for the agentic era — a unified, production-scale telemetry layer powered by Genie Code, our AI assistant, spanning every AI offering on the platform: Model Serving, AI Search, AI Gateway, Agent Bricks and Apps. Genie Code doesn't just alert you that something's wrong — it delivers instant root-cause analysis, pinpoints fixes for new error patterns and runs agentic operations on demand. Walk away knowing how to keep your application running smoothly in production — and how to get to the bottom of any issue in seconds when it doesn't.

Session Speakers

IMAGE COMING SOON

Anshul Gupta

/Sr. Staff Software Engineer
Databricks

IMAGE COMING SOON

Mingyu Li

/Staff Software Engineer
databricks

Full Summary

Why AI observability needs its own agent: inside Databricks' vision for Genie ZeroOps

AI systems have outgrown dashboard-driven troubleshooting. As models, agents, and apps spread across production, stitching metrics, logs, and traces by hand no longer scales. The session outlines why observability must become agentic and how Databricks is building toward Genie ZeroOps, where an agent investigates, explains, and can mitigate issues using the full system context unified on the Databricks Platform. For teams evaluating where AI workloads should live, the Databricks Platform explains how a unified foundation differs from stitching together point solutions. The core claim is simple: model intelligence is not the bottleneck anymore, context is. An observability agent that sees code, data, model lineage, and runtime telemetry can close the debugging loop far faster than humans jumping between tabs.

Observability is a context problem, not a tooling gap

The speakers frame observability across four dimensions: operational health, cost, quality, and governance. Modern models can reason well enough to identify and fix many issues, but only if they are given the right context. Metrics and logs alone hint at symptoms, not causes.

Databricks centralizes the ingredients an agent needs to reason about production behavior: the data, the models and apps, the code that produced them, and the telemetry they emit. Genie Code sits on top of that context to reach root causes rather than plausible guesses, which is the foundation for agentic observability.

A shared telemetry plane in Unity Catalog

Operational telemetry across Databricks Apps, serving, and agents lands in Unity Catalog: out‑of‑the‑box metrics, system and deployment logs, traces backed by MLflow tracing, inference tables for request payloads, and CRUD events. Genie Code uses these governed tables directly, so investigations start with full awareness of the endpoint, lineage, and runtime signals.

Recent improvements include near‑instant inference tables, sampling controls to manage cost, finer grained metrics across serving, AI search, gateway, and agents, and persistent logs across endpoint updates. Scale numbers underscore the maturity of the stack: more than 9,000 customers, over 170 million time series per day, more than 5 million traces, and four nines of availability.

From dashboards to proactive Genie ZeroOps workflows

Genie ZeroOps moves from reactive UI hunts to continuous monitoring and automated response. Watchers track SLAs, latency spikes, error rates, deployments, and periodic checks. When something triggers, Genie Code investigates, records the reasoning steps, and produces a root‑cause summary with suggested fixes. Depending on action policies, the agent can notify an operator or execute mitigations such as a canary rollback to a previous model version or rate limiting a single noisy client.

Skills extend investigations with organization‑specific knowledge. Teams can register custom skills in Unity Catalog so the agent speaks the same language as their code and services. Action policies act as guardrails that define which automated actions are allowed for each alert type, and which are explicitly forbidden.

One agent across Databricks Apps, agents, serving, and search

Production AI usually spans chains of components. An app calls an agent, which calls a serving endpoint that depends on an embedding model behind a search index. Because all of these emit telemetry into Unity Catalog, Databricks can construct cross‑request lineage and let Genie Code trace failures end to end. The same approach generalizes beyond uptime to cost budgets, quality thresholds, and governance signals, since those data also live in Unity Catalog.

What changes when the platform holds the whole picture

Agentic observability becomes practical when the agent can see the entire system, not just its symptoms. Centralizing telemetry and context in Unity Catalog enables Genie Code and Genie ZeroOps to reduce on‑call toil, shorten time to root cause, and pave a credible path toward hands‑off operations for AI workloads.

FAQ

Genie Code is the reasoning layer that investigates issues using metrics, logs, traces, inference tables, code, and lineage. Genie ZeroOps is the broader operational model that orchestrates watchers, skills, and action policies to detect, investigate, and, when permitted, mitigate issues proactively.

Out‑of‑the‑box metrics, logs, traces, and inference tables span model serving, AI search, AI Gateway, agents, and Databricks Apps. Genie Code can be invoked directly from endpoint pages in production today, with investigations grounded in Unity Catalog.

No. Action policies define guardrails per alert type. Teams can permit targeted mitigations like rolling back to a previous model version or limiting a single client, while disallowing destructive actions such as deleting an endpoint or scaling it to zero.

Yes. Custom metrics, logs, and traces from your code can be exported into Unity Catalog alongside built‑in telemetry. Teams can also register custom skills in Unity Catalog so Genie Code incorporates organization‑specific reasoning during investigations.

Databricks Apps, agents, serving endpoints, and search indexes all emit telemetry into Unity Catalog. Databricks uses that shared plane to build cross‑request lineage, which lets Genie Code trace a failure from the app layer down to an embedding model without manual context stitching.