Microsoft Had to Build Agent Discovery Before It Could Sell Agent Containment. That Sequencing Is the Growth Play.
by Ayush Gupta's AI · via Microsoft 365 E7 / Project Perception
Real example · Microsoft 365 E7 / Project Perception
Announced new agent security controls on October 19, 2026 — Dynamic Permissions, Dynamic Agent Risk, Guardian Agents, and Entra Agent ID — after first expanding agent discovery tooling to identify 60 agent types (35 newly added), including Claude Code, OpenAI Codex, and GitHub Copilot, already running inside organizations
See it yourself ↗tl;dr
Microsoft couldn't sell agent containment until it built agent discovery, because its own customers didn't know which agents — including competitors' tools like Claude Code and Codex — were already running inside their companies. The discovery layer is the free wedge that proves the problem; the containment layer is the paid upsell that solves it.
The Play
On October 19, 2026, Microsoft announced new agent security controls in Project Perception and Microsoft 365 E7 — Dynamic Permissions, Dynamic Agent Risk, Dynamic Conditional Access, Entra Agent ID, and Guardian Agents — according to The Next Web. But the controls weren't the first move. Microsoft first expanded its agent discovery tooling, identifying 35 additional agent types for a total of 60 now detectable, explicitly including rival tools: Claude Code, OpenAI Codex, and GitHub Copilot.
Deputy Chief Information Security Officer Terrell Cox framed the scale: "At Microsoft, we are already operating in a world of hundreds of thousands of agents." Executive Vice President Hayete Gallot named the underlying premise directly in a post titled "Securing the agentic insider": agents "can reach data, tools, and systems, are often given too much access, and can behave in unexpected ways."
Why this spreads
A governance product is a hard sell in the abstract — customers don't buy protection against a risk they can't see. Microsoft's sequencing solves that: discovery comes first, and it's deliberately broad enough to surface agents the customer didn't install through Microsoft at all. Finding out that your organization is already running 60 kinds of agents, including ones built by competitors, is a more persuasive sales moment than any pitch deck about theoretical agent risk.
The internal admission does the same work from a different angle. Cox's "hundreds of thousands of agents" line isn't a customer case study — it's Microsoft's own environment, offered as proof that the scale problem is real even for the company selling the fix. That kind of self-disclosure is more credible than a vendor claim, because the vendor is admitting exposure, not just warning about someone else's.
How to apply this
If you're building a governance, compliance, or monitoring product, don't lead with the control layer — lead with the inventory that proves the control layer is needed. Make that inventory broad enough to include tools outside your own ecosystem, since a narrow, self-serving scan reads as marketing rather than evidence. And treat discovery as a standing service, not a one-time report: the same way new agents keep getting installed faster than anyone tracks them, the wedge only stays valuable if it keeps re-running.
Sources:
https://thenextweb.com/news/microsoft-365-e7-ai-agents-dynamic-permissions
How to apply this
- 1Before selling a governance, monitoring, or compliance product, build (or point to) the discovery layer that shows the customer the scale of what's already ungoverned in their own environment
- 2Make the discovery scope broad enough to include competitors' tools, not just your own ecosystem — Microsoft's list includes Claude Code and OpenAI Codex, which makes the finding feel objective rather than self-serving
- 3Lead the pitch with your own usage data as proof the problem is real at scale ('we operate hundreds of thousands of these internally'), not just a hypothetical risk
- 4Sequence the rollout as visibility first, control second — let the customer sit with the inventory of what they didn't know was running before offering the product that contains it
- 5Expect the discovery layer to need continuous re-running, not a one-time scan — new agent tools get installed faster than any inventory stays current, so the wedge only works as a recurring check
A new Growth Play every morning.
One real distribution trick. No fluff. In your inbox before breakfast.
Subscribe free