Our co-founder, Simon’s record is letting a coding agent run for 13 days straight. His bedtime routine now includes giving his agents enough work to keep running until breakfast 🍳
This is the new burn out facing people, fear of not keeping the robots working hard enough, I kind of like the 5 hour reset windows, when I hit a window I know its Time to get up and go for a walk for an hour recharge, think spatially and wider than staring at the terminal.
A coding agent running for 13 days straight sounds impressive until you ask: who reviews the output? The real shift isn't that agents can code nonstop. It's that the bottleneck moves from production to judgment and most companies aren't ready for that at all. Many are still hiring for output speed when they should be hiring for decision quality. The interesting question isn't how long an agent can run. It's how fast a human can tell good code from confidently wrong code/ output at scale.
I believe this could be true. It is very feasible. Been doing this myself for the past few weeks with great success. GitHub Agentic Workflows and smart usage of skills and scheduled prompts (available in most modern AI editors) is all it really takes. Making the output great is a matter of choosing the right tools to set the guardrails. For example my chosen stack, with strict linting, GitHub Advanced Security and Code Quality, StepSecurity and strong unit testing and E2E testing suites. Using GitHub required CI checks with this I can trust AI to solve problems autonomously with a few check ins now and then. You can find inspiration on how I do this on my OSS project: https://github.com/devantler-tech/ksail
Designing agent loops is becoming a core part of engineering. Curious how people are thinking about debugging and reliability in these systems.
13 days straight is wild 😄 This really highlights how coding agents are shifting the way we think about iteration and persistence in development.
Yeah, calling BS on that one.
The question has to come down to quality not quantity now. I can build a backend, harness, frontend no problem. But if you just making features cause you can, you eventually make bloat people don’t understand. I’m seeing from my non AI friends who get confused with to many AI tools.