R093 · Infrastructure for agents at scale › Agent session hosting and control planes

kagent 1.0: one sandbox per conversation, pinned to its prompt and model

Read the original

agent-hostingkubernetessandboxingdurable-execution

Source: kagent-dev/kagent README (Apache-2.0, CNCF sandbox project originally from Solo.io); 1.0 GA announced November 2026. Surfaced by André Lindenberg.

What changed

Before 1.0, kagent ran each agent on Kubernetes like a web service: a long-lived Deployment that served every conversation. In 1.0, agents run on a Substrate layer that manages their lifecycle (suspension, resumption, snapshots), and, per the post, each conversation gets its own sandbox pinned to the prompt and model it started with. A prompt change therefore reaches only conversations opened after it.

Building blocks (from the README)

Why it matters

Tradeoffs

Review checklist

  1. Can a prompt or model change reach a conversation mid-task in our system?

  2. How do we force an urgent fix onto running conversations?

  3. Is each agent run isolated from other users' runs?

  4. Can we fork a conversation from a checkpoint to debug it?

R084 (always-on agent primitives), R035 (durable execution), R074 (policy on every mutation), R021.

Connections

See R093 on the knowledge map

Part of the Research notes on nowucca.com. Figures and claims are the source's own unless the summary says otherwise.