It started with an experiment. We pointed a coding agent at a flattened snapshot of a customer's database to see if it could answer real questions about their business. It could. It could also take actions, build internal tools, and automate the small tasks that quietly eat hundreds of hours every week.

We started Nightshift in May 2025 because we kept seeing the same pattern: AI that looked brilliant in a demo and broke down in production. We assumed the bottleneck was cost and inference, so we spent months chasing cheap GPUs and optimizing token throughput.

When that didn't close the gap, we decided the harness and application layer wasn't strong enough. We spent another few months (and more money then I'd like to mention) building more sophisticated orchestration, more app features, more scaffolding around the model.

Yet, more failed user experiences. More demos that didn't scale. More agents that couldn't do the work.

Then, as a last resort, we stripped all of that away. We connected every piece of data we could directly into Claude and just asked it to do the work. It exceeded every expectation we had.

The models weren't the bottleneck. The query loop wasn't the bottleneck. The infrastructure between the model and the work was. The lesson was simple that I will never forget it: stop building around the model you have today. Build for the model six months from now. Give it secure, sandboxed access to real systems, make the infrastructure consumable, and get out of the way. That's Nightshift.