Do you trust your autonomous agents enough to be accountable for their actions and outcomes?

Discussions
,
1

I’ve been trying to get my head around the idea that my toaster cannot be held accountable for burning my bread and setting the house on fire.

Okay so quite an extreme example but I keep coming back to this when I think about handing all testing over to an autonomous tool.

Here is an example from one tool provider “You set what good means. You make the release call. Studio does the work in between.” and I’m thinking well “good” is quite a big thing here, can I just let it run and still be accountable for the testing. Others may even be offering full test flows with testers out of the loop.

Is anyone using autonomous agents yet, can you provide some feedback from this angle in particular the accountability question but also if you were asked could you comfortably explain its intent, assumptions and risks associated with its actions? Any insight though would be of value, what works well, what doesn’t for example.

What about others, perhaps your management is pushing the idea of autonomous test agents forward, what sort of control and interaction levels do you feel you would need to take accountability?

Is anyone else having these discussions?

I also wonder if my own thoughts carry a bit of hubris on my testing ability, self preservation of my role or even just a why should robots get all the fun angle.

For those interested, I have written up some related more detailed thoughts here but from a broader development angle, touching on areas like the value of tacit knowledge that the tools do not have and in particular looking at this from human distance to tool aspect. Maybe that is also of use to testers in addition to the discussion here.

2

i don’t think the agent can be accountable. the release owner still is.

so the practical test is whether that person can reconstruct the call without rerunning the whole job: what goal the agent was given, which assumptions it made, what paths it exercised, what it skipped, and where a human accepted the remaining risk.

if the run can’t answer those questions, “you set what good means” is just accountability handed back after the work became opaque.

3

Dwayne nailed the frame: the agent can’t be accountable, the release owner is.

What I’d add: you can only be accountable for what you can inspect. Accountability is a property of the evidence, not the automation level — a fully autonomous run you can reconstruct is more accountable than a semi-manual one you can’t.

So to your intent/assumptions/risks question: you can only answer it if the run externalizes those — what it was told “good” means, which paths it exercised, what it skipped, and where a human accepted the risk. If that stays inside the tool, “you set what good means” is just signing off on a black box.

Trust the run you can re-read, not the one that asks you to trust it.