Guardrails, not guesswork
Singh Labs is me, Manpreet Singh, and a set of small tools that all answer the same question from different angles: an AI agent is about to change something — should it be allowed to?
Every team is now handing commit access to something that writes faster than anyone can review. The tooling around that has become very good at making agents capable and stayed almost silent on making them accountable.
The gap is not model quality. It is authority. Nobody asks whether they trust a new engineer in the abstract — they give them access to some things and not others, and they write it down. Almost nobody has written it down for the agents already committing to their repositories.
So each tool here is the same shape: a written rule, enforced by a program, leaving a record a person can argue with afterwards.
A safety tool that quietly relaxes is worse than none, because you stop watching it. So none of these tools will ever widen their own scope, guess at a version number, write a secret, or make a failing test pass. When one does not know what to do, it says so and stops.
I write up most of what I build on Medium — usually the reasoning behind a design rather than a tutorial. Occasional notes go out by email when something ships; the form is on the home page.
I take a small amount of contract work on agent infrastructure — approval gates, CI automation, and the boring safety layer around all of it. Get in touch.