Peter Lee

Yeah, I can give a real-world example that might help people here.

I mean, I think if you were to Google

where to get your hair cut.

For example, you might get a list of 100 different barbers in your local area.

If you use generative AI to ask where you should get a haircut in Canberra, you might get a more sophisticated answer, and you might then ask it further questions about the style that you wanted.

or how quickly you wanted it done if you ask an agent to sort your haircut out for you then the agent would have the autonomy and the ability to go away and book that appointment probably book your taxi there pay for the haircut as well so it's got the ability to conduct actions on your behalf all you have to do is get in the taxi and get there sit in the chair with the barber

Yeah, agents have got a level of autonomy that we haven't seen before from technology.

Therefore, you need to think carefully about the guardrails and permissions you give the agent so that you can be confident that what they end up doing and the actions they take are safe and secure.

Recently, for example, a couple of weeks ago Microsoft released a runtime governance model.

So this is real-time governance which sits across agents as they operate as a technology solution and allows you to spot when agents aren't performing as you expect them to.

If they're starting to drift, which means they're changing course, they're doing things that you're not expecting them to do,

And it allows you to monitor that in real time or near real time and then act upon it.

So if you need to kill the agent, you know when to do that.

What this relies upon as well is a different approach to policies.

I've started advocating for agent resumes or agent models.

We need to be a bit careful about anthropomorphizing this technology.

But I think people do find it quite useful to think about agents as digital workers sometimes.

And so giving that digital worker, the agent, a job description or a resume can really help people understand, especially non-technical people, to understand what that agent can and can't do and what it should or shouldn't be doing.

And that allows you to track its performance a bit like you would with a human worker.

And if necessary, you know, fire it.