# Agents need identities, sandboxes and auditable controls

Markdown copy of https://realalpha.app/s/agents-need-identities-sandboxes-and-auditable-controls/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Satya Nadella](https://realalpha.app/builders/satya-nadella/) · Satya Nadella is chairman and CEO of Microsoft. · Led Microsoft's cloud infrastructure and services transformation, and has been CEO since February 2014.
- Host: [Reid Hoffman](https://realalpha.app/people/reid-hoffman/) · Co-founder, Manas AI · Co-founded LinkedIn; one of OpenAI's first investors; Greylock partner.
- Show: Possible
- Episode: [Satya Nadella: AI Is the Future of the Firm](https://realalpha.app/v/BKx0Dp8y-6g/), Jun 5, 2026
- Moment: 16:14 to 18:51 of the episode (3 min), playing on https://realalpha.app/s/agents-need-identities-sandboxes-and-auditable-controls/; the episode from there: https://realalpha.app/v/BKx0Dp8y-6g/?t=974

## Quote

> “I first need to know what are these agents? What are they doing? **What are their reasoning traces**?”
> — [Satya Nadella](https://realalpha.app/builders/satya-nadella/) at [16:41](https://realalpha.app/v/BKx0Dp8y-6g/?t=1001) · Possible

## Takeaways

1. Maintain an inventory of agents and inspect their reasoning traces.
2. Give agents identities, sandboxes and policies for tool access.
3. Use runtime assertions to keep long-running agents within boundaries.

## About this segment

Satya Nadella says companies need inventories, identities, sandboxes and policy controls for agents. Runtime assertions and audit trails can help keep long-running agents governable.

From “Satya Nadella: AI Is the Future of the Firm”, Jun 5, 2026. [Watch the full episode](https://realalpha.app/v/BKx0Dp8y-6g/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[16:18](https://realalpha.app/v/BKx0Dp8y-6g/?t=978) It's a great point. So in fact, to your point, one is the experience layer. Clearly, that's important. And another one which we referenced earlier was, hey, we need this hill climbing machine as a concept that needs to be instantiated. But the third is the manageability of it and the security of it, starting with observability, right? One of the first things we did even, and we talked about it even yesterday, is something called Agent 365.

[16:40](https://realalpha.app/v/BKx0Dp8y-6g/?t=1000) I need to know, I need to have an inventory. I said, oh, there may be 20 million agents at Microsoft. I first need to know what are these agents? What are they doing? What are their reasoning traces? they need to be fully inspectable, fully auditable. And by the way, when the agents also have this other attribute, they are running, they can generate code and execute. So you need the environment in which they're executing code with maybe even file system access, network access to be think that is governed by policy. So therefore, you now need to give them identities, you need to give them sandboxes, then you need to set policies to govern them.

[17:20](https://realalpha.app/v/BKx0Dp8y-6g/?t=1040) And so we built this entire thing called Agent 365, which really has, you know, we have extended Entra so that you have an identity. We extended Defender so that you have security. extended purview so that you can even label the data that it's getting created automatically so that you can have data protection. So, I think security, containment, manageability, observability is the way we will have confidence around these agents. The one other thing I'd say, Reid, also was salient in our developer conference was how important it is when you're building these long-running agents, right? Like if you think about, right, we have always had in programming languages, you know, programming models for understanding the, you know, the verifiability of a program, right? And it's at runtime. So one of the attributes of these long-running agents

[18:14](https://realalpha.app/v/BKx0Dp8y-6g/?t=1094) that we added to Foundry was something called asserts. So this allows us to assert what are the boundaries, right? So instead of talking about guardrails as a thing that is sort of just a classifier of some form, you really need to now have the ability during execution to really have that execution path not go off rails. And so there's a lot of engineering sophistication now that's emerging as we build out the platform, the runtime, the security layer, the management