AI Agents Deserve the Same Accountability as Employees
Agent accountability is quickly becoming a business problem, not just an IT one. AI agents now work across finance, operations, sales and product teams. Leaders need a way to govern them that everyone can understand.
Gravitee’s Tim Tyler examines that challenge with Alan Shimel at KubeCon + CloudNativeCon in Salt Lake City. Tyler is a senior product marketing manager at Gravitee.
A Framework for the Digital Workforce
Tyler convenes the Agent Accountability Framework, an open framework sponsored by Gravitee. It grew out of conversations at industry events.
He kept hearing that technical protocols and standards were not enough. Business leaders wanted something methodology and technology agnostic.
His answer borrows from a century of human workforce governance. Instead of seeing AI only as software, treat agents as a digital workforce. The same principles that govern employees can apply to them.
Identity, Sponsor, Role and Authority
The framework rests on six principles. The first two cover assignment. Every agent needs its own identity rather than impersonating the person who calls it.
Each agent also needs a sponsor. That sponsor is a named person, not a team or a role. A named sponsor creates a clear line of accountability that runs up to the board.
The next two principles cover alignment. Role defines what success looks like, much like a job description. Authority defines what the agent can access and do to perform that role.
Oversight and Recourse at Machine Speed
Oversight is the fifth principle. It shows what an agent did, when and how, and whether it stayed in bounds.
Recourse is the sixth. When an agent strays, the organization must correct it. That might mean adjusting its role or blocking access to a database.
Timing is the big difference from human employees. A quarterly review comes far too late for an agent making thousands of decisions. Agent accountability has to work at machine speed.
The principles, controls and a GitHub repository are available at agentaccountabilityframework.org.
Transcript
Hey everyone, welcome back. You know what? This is it.
Last interview for KubeCon here in Salt Lake City 2026. I apologize for my voice, but it's just not here today. Luckily, Tim here is going to do all the talking on his interview, thank goodness.
I want to introduce you to Tim Tyler. Tim, it's great to have you on here. Thank you.
Oh, yeah. It's great to be here. So Tim, tell the folks about yourself.
Well, I'm Tim Tyler. I'm the convener of the Agent Accountability Framework. That's an open source business framework around agent accountability.
And I'm also a senior product marketing manager at Gravity. But I'm here talking to folks about agent accountability. Absolutely.
That's where you are today. Give us a little history. Yeah.
0, back when we were- Me too ... dealing with XML, XSLT- Yep ... through the DLL phase, all the way through the cloud native.
And now as we transition into agents, it's a whole new exciting world. It really is exciting. Talking all week about this, I feel lucky to be able to be here in the industry while all of this great stuff is going on.
So you're with Gravity. There are some companies, Tim, where they're good community players. Yes.
And they hire people knowing that those people are going to give a significant portion of their work time to open source projects, of which there are many, 240, 70, whatever it is. Yes. CNCI.
That's kind of the role of Gravity, or is this totally outside of the Gravity realm? No, this was sponsored by Gravity to convene this framework. That's great.
I noticed as I was going out to shows and talking to folks, that there was a need for a business framework, not just the technical protocols and standards, looking for something that was methodology and technology agnostic so that business folks could understand. So I started talking about it on the stage, making a metaphor between what we've learned over the last century of the human workforce governance. Why don't we stop looking at AI just as software?
What if we also look at it as a digital workforce that we apply those same principles to, and came up with this framework. Gravity said, "Yes, we like that. " People are really interested in the talks that I'm doing.
" That's because we've all seen that agents are not, and AI in general, is not just an IT problem. It's not just an IT promise. It is happening across the enterprise.
We have folks in the COO, the CFO, the CRO, the CPO. We have folks across all those organizations that are needing to find out how to work with AI, how to work with agents, how to replicate and expand and bring superpowers to the workforce that they have now. And so the framework was built around, let's look at it as a workforce, but keep it compatible and adjacent to the technological methodology that are all these other standards and protocols and frameworks that enable that to happen.
Okay. So from what I understand, there are six foundational principles. We call them principles, yeah.
Okay. Six principles. Let's go through them.
All right. So the six principles are broken down into three easy domains. So we've got the assign domain, the alignment domain, and the account domain.
In assignment, we are looking at, very simply, identity and sponsor. So what does that mean? Every agent should have its own identity.
If you are Al, the employee, and I'm Tim, the agent, we are separate identities. When you call upon me, I shouldn't just become Al. I shouldn't impersonate Al.
I should actually be Tim that has access into certain things within the organization, and I should then be delegated to run the task that you're asking of me. On the other hand, with sponsor, we need to have somebody who is accountable, who is responsible for that agent. And I like to talk about the line of accountability.
Once you have a specific person name, not a team, not a role, but a specific person, that specific person has already a built-in line of accountability going up through out through the board and into what the company has an accountability for what that agent does. And then we go to the alignment domain. The alignment domain is where we set the definition of what that agent is.
So we have role is the third principle, and authority is the fourth principle. And we have on role is a definition of success. Just like when you start a job, you got a new manager, that's your sponsor.
They tell you what you're supposed to do, how they expect you to do it. And so that's role, definition of success. With the authority, you also have what can you access, what can you do in order to perform the role, right?
So now we've got definition. But now we need to hold the agent accountable to that definition. So the sponsor wants to be able to see, are they behaving in the authority definition, and are they being successful at performing the role?
And so we need oversight. That's the fifth principle. So with oversight, that's what we often call on the technology side of this observability.
We want to know exactly what happened, when did it happen, how did it happen, and are they outside of the bounds of that role and authority definition? Because if they are outside of that, the last and sixth principle is recourse. What are you going to do to correct that?
So let me take it back to an employee metaphor. So Al, you're my boss. I go to work in your organization.
You got some oversight. You're noticing that I'm doing some things that are not quite to what you defined as my job description. Well, you've got some options.
You can have a meeting with me, have a one-on-one, coach me, allow me to become better at what I'm doing. In an agent context, you're just going to adjust the role. But also it might be that you need to take some action to say, "Tim, you're messing around with this database.
You shouldn't be. We're going to lock you out of that. " Well, we should also do that with agents, right?
We should have recourse. If they're starting to access things, we need to put in those blocks to keep them from accessing things they shouldn't. And so recourse and oversight, unlike a human, we might have that one-on-one quarterly, annually.
By the time you do that with an agent, it's 40,000, 400,000 decisions too late. So we need to be at machine speed, and so we need to figure out how do we actually perform oversight and recourse at machine speed, and that's when you get into more of the methodology. But the framework says you need to have this.
org, we actually have all the 45 different listed controls under those principles to be able to test and audit to see how do you actually build a proper role? How do you actually implement proper recourse? But again, in an agnostic definition.
Love it. Good stuff. Tim, for folks following along at home, where can they get more information?
Yeah. org. I know it's a little bit of a mouthful, but we try to make it as clear as possible for folks.
Descriptive. Yeah. org.
I don't know, but that would be pretty tricky. That's going to be pretty penny. Three letter domains are not cheap.
That would be pricey. Yeah. Very.
org and all those principles I was talking about, the controls and some content around the talks that I'm doing, the talks that others are doing around the world, and the participants in the framework- All there ... it's all on that hub. As well as the GitHub repo is there as well.
Cool. Hey, thank you for what you do. It's people like you and the maintainers, a lot of unsung heroes that keep this whole thing spinning.
So thank you very much. It's been a pleasure. Thank you for having me talk today.
I'm thanking you for talking today, too. Hey, that's it. We're wrapping.
My voice is gone. We'll see you next at AWS re:Invent. But for KubeCon, thanks for joining this Techstrong TV.