Agents Ask. Platforms Decide. (Why Both Layers Matter)

[Adeel Ahmad (adeelahmad)]

AGT solves the hard runtime problem: "is this action allowed at call time?" That's the enforcement kernel every platform needs.

But there's an upstream half that AGT doesn't touch: "what is the agent asking for before it runs?"

Today, that's invisible. An agent's model, tools, network access, timeouts — all buried in code or config files scattered across environments. No platform can read it, no security team can review it before deploy, no CI can gate on it.

I built agentrc to fix that: an agent declares its needs in an artifact (like a Dockerfile for agents). The platform reads the artifact — never the code — and AGT enforces the decision at runtime.

Same agent, any language, any framework. One artifact, any substrate. Platform sees it coming.

If this fits into AGT's direction, I'd rather see it absorbed into one standard than fork the space. The spec is open. The CLI is reference tooling.

https://agentrc.ai

What do you think — is the artifact layer something you'd want to integrate?

[Eray Özer (ozereray)]

I think “Agents Ask. Platforms Decide.” is a useful mental model for production agent systems.

The part I'd add is that the platform decision should become a first-class runtime event.

If an agent requests an action and the platform allows it, blocks it, or escalates it, that decision itself becomes security evidence:

agent → requested action → policy/context → decision → execution result.

That gives you something much stronger than a dashboard showing what happened after the fact.

This is very close to the architecture we're exploring with Aegisora, where the policy decision happens before the action crosses the execution boundary.

The interesting question for me is whether this decision receipt eventually becomes portable across frameworks rather than being tied to one governance implementation.