Agent Map & duplicate detection
625 agent records.
How many are actually different agents?
A raw agent count is the easiest number to inflate and the least useful to act on. The same agent promoted through sixteen environments looks like sixteen problems. The Agent Map groups an estate by what agents do, and duplicate detection separates a deployment pipeline doing its job from someone copying an agent because they could not find the original.
Two different questions get called "duplication". Is this the same agent in more than one place? is a deployment fact. Did somebody copy this agent instead of reusing it? is a governance problem. Reporting them as one number produces a red dashboard nobody can act on.
Three classifications, not one count
One platform definition present across several environments: dev, test, production. A pipeline copy or a vendor-shipped template. It inflates a raw count and is exactly what healthy ALM looks like. Reported, and excluded from the duplication signal.
The same name and the same instruction length as another agent. Two teams maintaining two copies of one capability, drifting apart from the day the second was made. This is the number worth acting on.
The same name with different content. A candidate for review, presented as a candidate - not asserted as a confirmed duplicate.
The Agent Map
Several hundred agent records collapse into around twenty capability families - flight tracking, ticketing, scheduling, invoicing - each drawn as a circle you can open. Three levels, never more than about twenty objects on screen, because node-link diagrams past that stop being readable.
What we do not claim
Duplicate detection compares names, instruction size and deployment lineage. It does not compare meaning.
Two agents doing the same job under different names
Not detected, and we will not imply otherwise. Semantic comparison needs the agent's instructions and topic definitions - which the platform management APIs used for discovery do not return. "No exact copies found" means the implemented checks found no match. It does not mean the family is free of duplication.
Why that distinction is the point
A governance tool that overstates what it detects trains people to ignore it. Every classification here is presented as what it is - a fact, or a candidate - and absence is stated rather than filled in.
Figures are from a connected Microsoft Copilot Studio and Power Platform estate at a design partner and are shown as an illustration of what discovery surfaces; your estate will differ. Scores and maps in product materials are labelled Sample / Illustrative.
Questions
What is AI agent sprawl?
Agent sprawl is the accumulation of AI agents faster than an organisation can inventory, own or retire them. It is made worse by raw record counts that treat one agent deployed across several environments as several agents, which inflates the apparent problem and hides the real one.
How do you tell a duplicate agent from a deployment copy?
By lineage and content. One platform definition appearing across dev, test and production environments is a replica - a pipeline copy, and normal. The same name with the same instruction length appearing as a separate definition is an exact copy, meaning someone duplicated the agent. The same name with different content is a variant, flagged as a candidate for review rather than a confirmed duplicate.
Can Govern360 detect two agents doing the same job under different names?
No. That requires comparing instructions and topic definitions, which the platform management APIs used for discovery do not expose. Govern360 states this limit rather than implying broader detection - no exact copies found means the implemented checks found no match, not that no duplication exists.
What is the Agent Map?
The Agent Map groups agents by capability rather than by the tool that built them, collapsing several hundred records into around twenty families. Families can be coloured by exposure, ownership, OWASP Agentic Top 10 conditions, or exact-copy density, so triage starts with whichever question is being asked.