THE SIGNAL IN ONE SENTENCE
Companies are building AI agents in every available corner: cloud platforms, customer systems, data warehouses, automation tools and internal AI studios. The awkward part arrives after the demo. Someone has to know which agents exist, who owns them, what they are allowed to do, how much they cost, whether anybody uses them and what happens when their behavior drifts. Dataiku announced a standalone product called Agent Management to assemble that scattered fleet into one inventory. Its product page names Microsoft Copilot Studio and Azure Foundry, Salesforce Agentforce, AWS Bedrock, Google Vertex, Databricks, Snowflake Cortex, n8n and Dataiku itself. That is eight platform families, depending on how one counts Microsoft's paired products. Dataiku says customers will select a framework, authenticate and let the system scan native APIs and log streams for agents. Each record is meant to carry an owner and purpose, then collect usage, cost, quality and business-value information. Higher-risk agents can be certified, assessed against named risks, retested on a schedule and attached to an exportable audit record. That is a sensible shape for a problem that is rapidly becoming inventory management with teeth. It is not available today. Dataiku says availability begins in October 2026. It also states that integration depth depends on what each platform exposes and that deeper integrations are still being developed. Those two sentences are the real story. A common dashboard cannot create a missing log, infer an owner from an abandoned prototype or observe an agent built outside the connected systems. It can normalize what it receives, flag what it understands and give humans one place to investigate. Dataiku also says the product is not an AI gateway. It sits above agents and tracks them rather than sitting in the execution path and controlling every request. The plain signal is that one inventory could make agent sprawl governable, but one screen is not the same as one complete truth. Before a company trusts the map, it needs connector-by-connector coverage, freshness, failure alerts and a reconciliation process for agents the scanners never see.
01
WHAT ACTUALLY CHANGED
Dataiku announced Agent Management as a standalone product for discovering, monitoring and governing AI agents across multiple platforms.
The company says the product will be available in October 2026 rather than being generally available at announcement time.
The product page names Microsoft Copilot Studio and Azure Foundry as supported sources.
It also names Salesforce Agentforce, AWS Bedrock, Google Vertex, Databricks, Snowflake Cortex, n8n and Dataiku.
Dataiku describes an automated scan that pulls agents from connected platforms into one inventory.
The planned inventory associates each agent with an owner and a stated purpose.
Dataiku says the system can track usage, cost, quality and business value against that purpose.
The product is meant to detect drift before users report a problem, according to the company.
Organizations will be able to certify agents that are cleared to run and assess them against named risks and mitigations.
Dataiku says tests can be repeated on a schedule and the resulting risk record can be exported for auditors.
The product connects to third-party platforms through native APIs and log streams.
Users are expected to select a framework and authenticate before Agent Management scans that environment.
Dataiku explicitly says integration depth depends on what each platform exposes.
The company says deeper integrations are still being developed.
Dataiku distinguishes Agent Management from an AI gateway because it observes agents from above rather than controlling traffic between agents and tools.
The product page does not publish pricing, connector-level field coverage, polling frequency, failure behavior, customer results or independent performance measurements.
02
WHY THIS MATTERS
An organization cannot govern an agent it does not know exists. Inventory is the first control, not the glamorous one.
Agent sprawl is harder than ordinary software inventory because one agent may call models, data stores, tools and other agents across several vendors.
A named owner creates a person who can answer why the agent exists, what data it uses and whether it should still be running.
A stated purpose gives monitoring a reference point. Without one, a dashboard can show activity but cannot say whether the activity is useful or inappropriate.
Cost tracking matters because agent workflows can loop, retry, call expensive models and trigger paid tools without a simple per-seat budget.
Usage matters for the opposite reason. A dormant agent can keep credentials and access long after its business case has disappeared.
Quality is not one universal number. A customer-support agent, coding agent and invoice agent need different tests, thresholds and consequences.
Business value is even harder to normalize. Time saved, errors avoided, revenue influenced and tasks completed are not interchangeable measures.
A common inventory can help security, legal, finance and business teams examine the same record instead of maintaining separate spreadsheets.
The connector limit is fundamental. If a vendor API omits prompts, tool calls, errors or identity details, the oversight layer cannot recover them through optimism.
Log schemas change. A connector that mapped every field last month can become incomplete after a platform update unless coverage is continuously tested.
Authentication can fail silently or expire. A dashboard needs to distinguish no risky activity from no current visibility.
Agents built with scripts, open-source frameworks or personal accounts may never appear in the eight named systems.
Because the product is not a gateway, discovery does not equal prevention. It may identify a problem without being able to stop the next action directly.
Scheduled certification creates a paper trail, but a certificate becomes stale when the model, prompt, tool permissions, data or business process changes.
The broader market is moving from building agents to accounting for them. The winning governance product will be the one that shows what it cannot see as clearly as what it can.
03
WHERE IT COULD HELP
- Require every agent record to include a human owner, business sponsor, purpose, users, data sources, tools, model, environment and retirement date.
- Show connector health and last successful synchronization beside every metric so stale visibility cannot masquerade as a quiet system.
- Publish a field-level coverage matrix for each platform, including identities, prompts, tool calls, costs, errors, outputs, policy events and deletions.
- Alert when a connector loses permission, stops receiving logs or returns fewer agents than its recent baseline.
- Reconcile the central inventory against cloud billing, identity systems, source repositories and network activity to find agents outside the official platforms.
- Assign a stable identifier that survives renaming, deployment changes and movement between development and production.
- Separate prototypes, tests, dormant agents and production agents so experimentation does not inherit production trust.
- Define quality tests by purpose and risk instead of applying one generic score to every agent.
- Record who approved an agent, which evidence they reviewed, the decision date, expiration date and conditions of approval.
- Invalidate certification automatically when the model, system prompt, tool list, permissions, connected data or evaluation suite changes materially.
- Track direct model cost, tool charges, infrastructure cost and human-review time rather than reporting only token spend.
- Compare measured business outcomes with the original purpose and retire agents that are expensive, unused or duplicative.
- Keep raw evidence links behind every normalized dashboard field so investigators can return to the source event.
- Create role-based views for engineering, security, legal, finance, audit and business owners without creating different facts.
- Define a manual containment route for high-risk agents because an oversight product outside the execution path may not stop them itself.
- Test every connector with known synthetic agents and events, then publish which expected records were received and which were missed.
- Export the inventory in a portable format so the governance record does not become locked inside the product monitoring vendor lock-in.
- Treat unknown and unavailable as explicit states. Never convert missing telemetry into a reassuring zero.
KEEP A HAND ON THE WHEEL
Agent Management is an announced standalone product with availability scheduled for October 2026. The current product page describes intended capabilities rather than published results from a generally available service. Dataiku names eight platform families, but that list does not prove equal depth, complete discovery or identical metrics across them. The company says integrations use native APIs and log streams and explicitly warns that depth depends on what each platform exposes. The product is not a gateway, so it does not necessarily sit in the path of an agent action or block it. Terms such as quality, value, drift and certification require customer-specific definitions and evidence. Dataiku has not published connector-by-connector event fields, API permissions, synchronization frequency, latency, identity resolution, deletion behavior, outage handling, pricing, benchmark results, missed-agent rates, false-alert rates or independent customer outcomes. Watch for a public coverage matrix, launch pricing, actual October availability, supported regions, evidence freshness indicators, export formats, customer case studies and a clear distinction between monitoring an agent and controlling it.
04
TERMS WORTH KEEPING
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on September 25, 2026.
PUBLICATION RECEIPT: Revision 1. Published September 25, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US