We're spending enormous effort making AI models smarter.
But are we spending enough time thinking about the infrastructure we're asking them to operate?
Most infrastructure was designed for experienced humans sitting at terminals. Commands such as:
š“šŗš“šµš¦š®š¤šµš š³š¦š“šµš¢š³šµ šÆšØšŖšÆš¹
make sense because the engineer brings the missing context.
They understand the system, its dependencies, the risks, and what might happen next.
An AI agent gets the command.
It doesn't automatically get the operational understanding behind it.
That's why I believe the architecture needs to change.
Instead of:
šš ā šš”šš„š„ ā šš§šš«šš¬šš«š®ššš®š«š
I think we need:
šš ā ššØšÆšš«š§šš šš©šš«ššš¢šØš§šš„ šš®š§šš¢š¦š ā šš§šš«šš¬šš«š®ššš®š«š
The AI determines šøš©š¢šµ should happen.
The runtime determines whether, and how, it can happen.
That runtime can provide things a shell command inherently doesn't:
ā¢ Structured resources and capabilities ā¢ Dependency context ā¢ Risk evaluation ā¢ Policy enforcement ā¢ Dry-run ā¢ Explain-before-execute ā¢ Approval boundaries ā¢ Auditability ā¢ Verified outcomes
Instead of teaching AI to generate every possible command, we can give it resources it can reason about.
Conceptually:
š“š·š¤://š¤šš¢šŖš®š“-š¢š±šŖ š³š¦š“šµš¢š³šµ
The syntax isn't the important part.
The important part is that the runtime understands the resource and the requested operation before execution occurs.
And when the model doesn't cover an operation?
It should say šš»ššš½š½š¼šæšš²š± rather than silently giving the AI unrestricted shell access. Exceptions can still exist, but they should remain governed, scoped, and audited.
This is one of the ideas I've been exploring with the open-source š„š²šš¼ššæš°š² š¦šµš²š¹š¹ (šæš²ššµ) project.
I don't believe the future of AI operations is about giving increasingly intelligent agents increasingly powerful terminals.
I think it's about creating a stronger boundary between šæš²š®šš¼š»š¶š»š“ š®š»š± š²š š²š°ššš¶š¼š».
AI can propose.
AI can investigate.
AI can plan.
But a governed runtime should decide what gets executed.
As AI moves from assistant to operator, do you think our existing infrastructure interfaces are ready or do we need to rethink the operational layer itself?