Agents started acting on their own. Someone had to keep them in check.

Aethis started in 2026 as a small side project. The original problem was simple: I wanted a reliable way to stop AI agents from overspending.

Reactive cost controls report a limit after it has been crossed. I wanted a budget enforced at execution time, so concurrent activity could not race past it. Prevention had to be part of the request path.

Then I watched an AI agent give someone access to an account that wasn’t theirs. The question was no longer just how much an agent could spend.

It became: what is this agent allowed to do?

Agents now use tools, APIs, databases and accounts to do work with real consequences. They need the same basic boundaries we expect for people: identity, permissions, spending limits, approvals and an audit trail.

Aethis puts those boundaries between an agent and the systems it can affect. It checks the concrete action before execution, brings in a person when needed, and records what was allowed or blocked, and why.

You should be able to run agents with the models, frameworks and providers that suit your application. Aethis is meant to govern across that stack, rather than become another platform you have to move into.

I’m building Aethis from Singapore, founder-led for now, because agent infrastructure is becoming consequential and governance cannot stay an afterthought. The product is early, and I’d like to hear from people building, evaluating or challenging this approach.