I already put this in the log, but whatever. Free will n all that, I'm double posting.

Been thinking about the future of SWE and agentic use in general (shocker, who working in SW hasn't?)

In my eyes: verification and governance.

My prediction on SWE directions: verification engineering becoming a massive discipline/the new standard for software engineering at large.

We're no longer hand crafting pieces of curated, artisanal code to try and solve problems.

Aside: people sometimes seem to forget solving problems was the point of the job in the end - writing code in and of itself was not the full job, but many people seem to have made it their personality or their measure of value/impact as an engineer, and are now struggling with their identity as a result, but that's a different thread.

I understand the tension here. Hand writing code gave many of us the level of scrutiny that allowed for an ongoing understanding and verification in terms of the code produced, but as the be all and end all of a given engineer's identity, it's now far too limiting.

We're now at the factory's production line, and our job is to assess which of the different experiments and approaches output by an agentic session best address a given problem, which is most extensible and/or maintainable for the future and its problems to address, and to provide a level of confidence and rigour to how we arrived at that decision.

How do we do that? With the necessary guardrails, assessments, and verification practices in place.

Zoom out, same problem appears at a bigger scale. Verification is governance at the scale of a PR; governance is verification at the scale of an organisation.

Governance.

Not the sexiest option by a mile, but it feels like the right route, because we're slowly (or quickly depending on your personal perceptions and p(doom)) offloading our cognitive capabilities to a subscription service.

Intelligence and aptitude on demand, where we need to pay a given value each month to be in with a fighting chance in the ongoing arms race of individual capability.

I'm convinced that, of all the AI safety risks we need to guard against, gradual disempowerment and over-reliance are far too easily overlooked in favour of misuse, for example. Misuse / CBRN applications are obviously terrifying, and real problems, but the insidiousness of disempowerment is the killer to me.

In X industry, Y organisation is so much further ahead than us! We need to use AI to catch up. We don't have time to spend dwelling on these decisions and problems - DecisionBot 9000, what's your suggestion here? Ok great, implement that.

Fast forward Z months/weeks/days later - I don't understand any of this. Where are we? Let's ask AI - it's best informed and capable to consume the information we have here, get its assessment and then we'll go with its recommendations.

At what point do we recognise the problem that's developed? How do we even begin to pull ourselves out of that hole? Do we even recognise we're boiling along with the other frogs?

We need governance in place to try and offset these situations at the earliest possible points. In practice: named human(s) owning each significant decision, with the ability to explain it without asking a model, and recording why a given recommendation was accepted, not just that it was.

I don't wanna be a dang frog.