I have been in software for over 20 years, and the last 2 have been the most insane emotional roller-coaster.
Backstory:
Early in my tenure at <current job>, before I was making use of agents, I caused a production regression in a legacy app while adding Python debugger support. This wasn’t caught in review or by the test suite. The irony of it being that the effort was supposed to help the less tenured members of the team (such as myself) — who lacked the tribal knowledge locked away in the heads of the taciturn veterans — work more confidently and cause fewer regressions.
Stories like that suck. I don’t want anyone to have to feel responsible for causing an outage, or imposing a burden on the person who has to wake up to an alarm at 2am to fix it; I have been in both boats, and I don’t want to be in either again.
Anyway, the lesson was not “if it ain’t broke, don’t fix it,” it was that good intentions are no substitute for sufficient context and executable proof — things that feel more necessary now than ever before.
Fast-forward to a few months ago: agents everywhere, cranking out code faster than anyone can reasonably review. Everybody’s trying to figure out how to “work with AI” in their own way, human PR review remains a necessary bottleneck, and I end up buying a subscription or two…
And I build shit, fast, because I can. I’ve been waiting for this for so long: describe a problem, and the solution appears — in a way that’s indistinguishable from magic — agents got me hooked, and gave me a new job that I’m literally paying to keep: using agents to fix the janky software that agents built for me. Sure, they write and run tests, but what does “all green” mean on a suite you didn’t review?
Free time evaporated due to context switching: every return to a side project began with reorienting, rebuilding context, preparing a sandboxed dev environment, and figuring out what of the last agent’s code was actually keepable, before doing any useful new work… and by then I’d be too tired to feel confident about anything.
Invariably I’d end up telling an agent to just go figure stuff out and build something (maybe after glazing over reading a massive plan doc) — which they are sometimes excellent at, don’t get me wrong — and I’d check in later. But then I’d have to review a bunch of code and reconstruct the mental model of the implementation after the fact, which is so exhausting…
… and corporate is encouraging people to work this way? Terrifying.
tl;dr: I waited a long time to try the flashy agent harnesses, was overwhelmed by their power, and ultimately felt unsatisfied and yet unable to escape their clutches…
So, I made a thing about it: nosedive.dev Now I keep all my ideas, notes, todos, bugs, specs, and a bunch of other stuff in a knowledge base in my personal “back-office” metarepo. I split work into small vertical “dives” in advance, using an agent with access to the full context, and can then use a fresh agent session to implement each slice in autopilot mode in a consistent way, while having them report progress and blockers back to the KB. Then I review and merge as many small, standalone, chunks of code as I have time for, and leave the rest packed for later.
I built this for myself (so, admittedly the onboarding docs are not great, but that’s what agents are for!), but now my team at work uses it every day too. Maybe it’ll be useful to you — as inspiration, if nothing else.
I can honestly say: building this tool — with agents, for agents — has been the most fun and rewarding software work I have ever done, and I would happily work on it full-time if I could.
My wish is that this software *revolution *we’re experiencing right now eventually makes life easier for all of us… and if it doesn’t? Well, I for one welcome our new agentic overlords. 🤖
Stay sane out there.