Dataiku Agent Management

Dataiku Agent Management, announced on 24 September 2026 at the company's Dataiku Succeed conference, answers a question most large organisations currently cannot: how many AI agents are we actually running, and which of them should worry us? It is sold as a standalone product rather than a Dataiku platform feature, and the bet behind it is that cross-platform coverage beats whatever agent console each vendor ships natively.

Discovery first. It inventories agents wherever they were built — AWS Bedrock, Databricks Agents, Google Vertex AI, Microsoft Copilot Studio, Azure AI Foundry, Salesforce Agentforce, Snowflake Cortex and Dataiku itself — and picks up anything custom through OpenTelemetry. For each agent found it identifies the structure underneath: which tools it calls, which models it runs on, and what it depends on.

Then measurement and risk. Agents are tracked against business KPIs as well as technical performance, so cost and quality are attributable per agent rather than pooled into one cloud bill. The product tiers agents by risk, tracks certification status for the high-risk ones, keeps risk documentation, runs scheduled testing, and maintains an audit trail for compliance teams. A plain-language query layer sits over the whole portfolio, so "which agents touch customer PII and have not been tested this quarter" is a question rather than a data-engineering task.

Dataiku's framing is that ordinary observability tooling was built for services with deterministic failure modes and misses how agents actually go wrong. CEO Florian Douetteau's line on the sprawl is blunt: "Nobody set out to build it this way. Teams built agents faster than anyone could count them."

Availability and pricing. Announced 24 September 2026, with general availability in October 2026. Access is priced per instance annually, with monitoring metered by the number of agents under management — enterprise procurement, not self-serve signup.