Keep it simple, keep it safe

When we sit down to a hard problem at Pendulum, we come back to the same two lines, Keep it simple. Keep it safe. Not because complexity is not required, but because a system almost always has more of it than the job needs.

Simple is better, always, when it comes to software. Fewer moving parts and fewer places to be wrong keep a surface you can still see and understand when the author is not in the room. We've inherited many systems over the years, and always give a quiet thank you to engineers who clearly favoured precise and simple procedures, over complex abstraction for edge case coverage.

AI changed who gets to start, which is more dangerous than it sounds. You can throw prompts at a window and watch a near one-shot of a system come back - routes, db schema, auth cycles, a dashboard that would have taken a fortnight. And everyone is doing it! Founders do it, internal teams do it, people who have never shipped a basic website before do it. It's a genuine technological marvel!

That is the real today payback for AI, and it is also how complexity arrives without anyone choosing it.

This is where the safety breaks down. The model does not know your risk and so it will invent it. It will solve all the problems through another queue, a second store, and a third agent if the prompt sounded ambitious. It will assume the boring parts that make software safe: who can see the data, what happens when a write fails, whether the secret stays on your box. Most of our rework from agent output is simplifying those assumptions, removing complexity so future agents and developers know what keep it simple, keep it safe means.

The 5% problem is still the 5% problem. The demo is cheap and production is the work.

This is not a lecture for juniors. It is the same rule for a studio, a founder pairing with Cursor, or someone who just got their first working app.

If you are using AI, follow the mantra. If you are directing agents, tell them to follow it.

Write it into the brief. Put it in the rules file. Say it before the first generate: prefer the smallest thing that works. Do not add a service, a graph, or a new store until the simple path is not enough. Do not ship a path that touches money, identity, or other people's data until a human has named the failure modes.

Agents match the appetite you give them making them the perfect development partner, if you know how to handle them. Ask for impressive and you get impressive, but ask for simple and safe and you get the ability to deliver both.

We have made this argument in other shapes. Own the architecture. Plan before you generate. Do not treat a multi-agent graph as a maturity badge. Stay on point. The common thread is restraint and knowing what you're doing.

Safe is the habit of refusing the extra moving part until it earns its keep - and keeping a person responsible for the parts that can hurt someone. Data ownership is defined, and you know who, where and how it's transformed to remain compliant. Reviewers that can say no when they should, because the decision scope is beyond that of your agents context.

AI compresses exploration, but it does not define what "done" means. We keep people on those boundaries so the application of keeping it simple results in keeping it safe.

The feed will keep celebrating one-shots, outlandish complex solutions, and many of them will be good. The ones that last will still be the ones that stayed simple, are easy to understand, or boring enough to trust.

Keep it simple. Keep it safe. Tell your agents the same.