Dataiku has launched Agent Management, a standalone product that inventories AI agents built on third-party platforms, alongside an expansion of Cobuild, its natural-language builder for AI projects. Fewer than one in five organizations keeps a complete and current inventory of AI systems, according to IBM research cited by the company. The gap matters because agent fleets grow faster than oversight, leaving compliance, cost and risk without a single owner.
How the shared inventory and Cobuild expansion work
Once connected to the platforms where agents run, Agent Management pulls them into one inventory. Connectors cover Salesforce Agentforce and agent services from Amazon Web Services, Microsoft and Google, plus agents built on Databricks, Snowflake and Dataiku, while custom environments report through OpenTelemetry. Each entry maps the tools and models an agent depends on, and the riskiest agents carry certification status with named risks. Tests on those agents rerun on a schedule, building an evidence trail for audits. The product will be generally available in October for an annual per-instance fee, with monitoring metered by the agent.
Cobuild became generally available in June as an agent that turns a business request into a governed Dataiku project without code, and the update extends that pattern. Cobuild Insights lets any employee query governed company data for what Dataiku describes as analyst-quality answers, then continue into deeper work inside Dataiku. Dataiku Headless brings Cobuild into Claude Code, Codex and Cursor, where developers can build or change pipelines, models and agents. An old Microsoft Excel workbook or Alteryx workflow can be converted into a governed project, and the tool can examine a live environment. Inside Agent Management, Cobuild recommends business-value metrics or diagnoses risk alerts, while the AI Catalog assembles vetted assets and semantic models into an answer or visual pipeline with full lineage.
Dataiku positions Agent Management above single-vendor monitoring tools, which it says tend to favor their own agents, and lets users ask portfolio questions in plain language, such as where risk sits or which agents cost more than they return. The pitch lands against survey data presented at the Succeed conference in New York in the Global AI Confessions Report: CIO Edition, 2026, based on 685 chief information officers at companies with more than $500 million in annual revenue in eight countries, polled by The Harris Poll on July 9-29. Nine in 10 said they were confident in complete tracking, yet 81% lacked complete oversight of agents built outside approved channels. Dataiku co-founder and chief technology officer Clement Stenac said technical and business teams often run on separate platforms without shared data or guardrails, a setup that a few pilots tolerate but that turns into a liability at scale.
What this means for companies using AI agents
For companies operating agents across several clouds and SaaS tools, a cross-platform inventory changes daily control: ownership, model and tool dependencies, certification and test history sit in one place instead of scattered consoles. Plain-language queries lower the barrier for risk and finance owners to find concentration points and loss-making automation without waiting for engineering reports. Business staff gain governed self-service through Cobuild Insights, while developers keep working inside Claude Code, Codex or Cursor through Dataiku Headless. A small firm gets a faster path from request to governed project, while a large bank or retailer gets the portfolio view that co-founder and chief executive Florian Douetteau described as missing when leaders can count servers but only guess at agents.
The announcement leaves several points to verify before purchase. Coverage depends on connectors and OpenTelemetry reporting, so teams should confirm which versions and configurations of Agentforce, Amazon Web Services, Microsoft, Google, Databricks and Snowflake are discovered automatically. Certification records, named risks and scheduled retests will need mapping to the buyer audit rules, including what evidence is stored and for how long. Business-value metrics recommended by Cobuild and cost-versus-return comparisons require agreement on definitions, or different units will read the same dashboard differently. The news alone does not solve agents built outside approved systems, the category where 81% report blind spots, nor does it remove the pressure noted by 84% of CIOs who say employees build faster than IT can govern.
The marker to watch is October, when Agent Management becomes generally available and early disclosures show which connectors, test cadences and per-agent metering terms ship in practice. A second marker is budget behavior: 72% of CIOs expect cuts or a freeze if this year performance targets are missed, and 76% see roles at risk if measurable business gains from AI do not appear by the end of 2027. Together those dates will show whether inventories turn into governed returns.
