Skip to content
refplat

Learn: The Agentic Platform

How the platform runs AI agents as governed workloads — bound so a compromised agent can’t cause harm, and made to earn trust rather than be given it. This is the platform’s most distinctive and most deliberately cautious area. What follows is a whole-picture tour plus focused deep dives.

Audience: platform engineers, and anyone wary of letting an LLM near production. It helps to have read Environment API (XAgent is its sibling), Identity (an agent is a subject in the same model), Observability, and Supply chain, but none is required.

Read in this order

Section titled “Read in this order”
  1. Orientation — “treat the agent like a powerful contractor you don’t fully trust.” The one idea, and a tour: what an agent is (an XAgent claim), how it’s bounded, and how it could earn more (the autonomy ladder), ending on the one real agent. Includes an honest “one agent, mostly early” status.
  2. Reference — the dense lookup: the XAgent claim and Composition, the GitOps control plane, the bounding machinery, the autonomy ladder, the triage copilot, the status ledger, and gotchas.

Go deep

Section titled “Go deep”
  • The XAgent runtime — the claim → slot Composition, the XEnvironment parallel, the GitOps control plane, Bedrock, and the lifecycle.
  • Bounding the agent: identity & safety — the scoped Pod Identity, the envelope, least privilege, the network lock, the data boundary, the kill-switch, and the admission gates.
  • Autonomy & evaluation — the graduated, eval-gated autonomy ladder (designed), the eval loop, and the forward-capture substrate that is built.
  • The triage copilot — the one live agent, end to end, plus owner-routing.

Then

Section titled “Then”