Infrastructure teams are already running AI agents against their cloud and on-prem environments. For every change an agent makes, teams still need to know who asked for it, who approved it, and where the record is.
For people, access is provisioned. There’s a named user in the identity provider, someone chose the roles, and there’s a record of who has what and why. Agents tend to pick up access along the way. It might be an API key pasted into a config, a developer’s own cloud credentials, or IAM rights to create roles and pass credentials to other agents. That’s what we call emergent access. Most teams have at least one agent running this way today.
This guide is based on CloudBolt’s webinar with The New Stack, “Provisioned access vs. emergent access: Tracing AI agents back to the humans behind them.”
An agent is just another client
To a control plane, an agent is a non-person entity making API calls, and control planes have governed those for years. The agent signs in as a user through the identity provider you already run, and it gets that user’s groups and environments with no broader access. Every write goes through the same approvals, quotas, and gates as a person’s request, and every action leaves the same records.
Access works in three tiers. Read is where every agent starts. Write is granted only when the agent asks, and it never exceeds what the user could do by hand. Approval applies to the actions you mark, such as production changes, destructive actions, and anything that grants access. For those, a different person approves or denies the request, and the decision is logged.
The five questions
Each question checks one part of how an agent got its access and what it can do with it. If you can’t answer one, that agent has emergent access.
- Who is it acting for? A named user in your identity provider, signed in through SSO. Not a shared bot account or a pasted key.
- What can it reach? That user’s groups and environments and nothing broader, with cloud credentials kept in the control plane.
- What needs a human? Reads by default, writes on request, and a different person’s approval for production and destructive actions.
- Where is the record? A request log line, an order, a job, and a resource history entry, all naming the same user and joined by ids.
- How do I turn it off? Disable the user in the identity provider, and every session they opened ends.
What changes when clouds run the work with their own agents
Cloud providers and integration tools are building agents into their own endpoints. AWS MCP Server, Amazon Bedrock AgentCore, Microsoft Foundry, Google Agent Runtime, and the Terraform MCP server all run agents inside your accounts. You send one request, the provider’s agents make hundreds of calls inside its boundary, and your control plane doesn’t see the individual calls.
When that happens, three things break.
- Completion. You get a report rather than a view of whether every agent finished.
- Attribution. Calls carry the role they ran under, not the order that asked for them, and sub-agents make it harder to trace.
- Consistency. Every cloud and tool ships its own agents, logs, and identities.
Five moves to start on now
These are worth starting now, whether provider-run agents reach your environment next quarter or next year.
- Templates are the contract. Keep templates clean, parameterized, and in git, so whatever runs the build only resolves to them.
- Policy before execution. Decide entitlements, quota, and approvals against the whole request up front, not call by call.
- A plan you can approve. For anything beyond a disposable environment, see what the build intends before it runs.
- One system of record. Issue a request id that travels with the work through every agent it spawns, and bring the evidence back to one place.
- Watch the endpoints. When a cloud describes a tool as an outcome (“provision an environment”) instead of an API call (“create a bucket”), that provider has started running the work with its own agents.
What’s in the PDF
The two-page PDF covers each of the five questions with what a good answer looks like and how to check it, plus a starting point for each of the five moves. It’s built for platform, infrastructure, and security teams reviewing the agents already in their environments.
Run the five questions against your own agents
Get CMP free
Download PDF
Joanne Chu
Senior Content Marketing Manager
Related Blogs
VMware Aria alternative for VCF 9: Aria vs CMP – Part 2
Part 2 of 3. Read Part 1: why companies are making the move | Read Part 3: testing VMware alternatives…