Table Of Contents
The Day I Left Claude Running Overnight
We've had a little competition at Wakeflow: Who can make AI work on it's own for the longest. Months back it was around 20min. Weeks ago it was a fair few hours...
This week I got to over 24h straight. I had Claude working on modernising a legacy codebase - totally on it's own. Without anymore input from me. I went to bed, I ran errands, I slept again, and it kept going — reading the codebase, making changes, running the tests, reading the failures, fixing them, and moving on to the next thing on the list.
So it now can do days of real, compounding work — refactors, tests, fixes, the unglamorous middle of a software project — with me checking in rather than holding its hand.
How it works
- Give it a goal, not a task. Instead of one instruction, I handed it a clear objective and let it break that down into its own work list. The agent decides what to do next, not me.
- Wire it into the real feedback loop. The key is letting it run the same things I'd run — the test suite, the linter, the build. When a change breaks something, the failure comes straight back to the agent and it fixes it. No human in the middle to relay errors.
- Let it loop. I set it up so that when one piece of work finishes, it picks up the next one automatically and keeps going until the list is empty. That's what turns "a clever reply" into "a night of work".
- Keep everything in version control. Every change is a commit. If a run goes sideways, I can see exactly what happened and roll back. That safety net is what makes leaving it alone overnight feel sane rather than terrifying.
What surprised me
What surprised me is the standard of the result. It is a really very accurate replica of the legacy codebase. It works in much the same way in terms of functionality, but the UI looks a million times better.
What was interesting to observe is: As it's working it comes up with more tasks that it didn't realise it needed in the original assessment and road map. This is very close to real software development where as you go you discover more detail and course correct. It was reassuring it does it like me, not with one anticipated mega master plan from the beginning that it then just executes.
The other surprise was emotional, honestly. There's something strange about closing the laptop and knowing work is still happening. It reframes what my job is: less typing, more deciding what's worth building and checking that it was built right.
Bigger picture
This matters far beyond a fun overnight experiment, because of what we're actually building it for.
One of the things we're working on at Wakeflow is buying legacy software firms — the kind run by a founder who's ready to retire, sitting on solid, profitable software that's been quietly running someone's business for fifteen years but hasn't had a real upgrade in a decade. We acquire those firms and modernise their tech stack with AI.
In practice that usually means taking an on-premise product — software that gets installed on a server in a cupboard somewhere and is a nightmare to update — and moving it onto the cloud. That one move makes the software dramatically easier and cheaper to operate. And once it's on the cloud and wired into a CI/CD workflow like the one above, we can do something the original product never could: ship improvements daily. New features, UI fixes, upgrades — pushed out continuously instead of once every couple of years on a CD-ROM.
Now connect that to the overnight experiment. An AI agent that can work for 24 hours straight inside that CI/CD loop is exactly what makes modernising decades-old codebases economically viable. The grind of understanding an old system, writing the missing tests, and steadily refactoring it — the work that used to make these projects too expensive to bother with — is the work agents are getting genuinely good at. This is how a small team takes on software that would once have needed a small army.
Get in touch
Got a legacy software firm in mind that would benefit from this? Maybe it's a founder you know who's thinking about retirement, or a product your business relies on that hasn't been updated in years and runs on a server nobody wants to touch. Let us know — modernising exactly these kinds of businesses is what we do.
And if you just want to talk about getting AI agents to do real, sustained work inside your own codebase, I'm always up for that conversation too. Reach out — I'd love to hear what you're sitting on.
