kagent 1.0 GA: Agent Substrate brings sandboxed AI agent execution to Kubernetes beyond pod-per-agent

bibryam · x · 2026-11-13

Solo.io announced kagent 1.0, an enterprise-ready distribution of Agent Substrate for running AI agents and sandboxes on any Kubernetes distribution including OpenShift.

The core argument: Kubernetes Pods were built for long-running microservices, but modern agent usage differs in two patterns:

Agent harness: a general loop where a model gets a goal, workspace, and skills/tools/CLIs and plans its own steps (Anthropic's definition of agents vs workflows), runs model-written code, keeps a workspace between turns, and mostly waits on models, tools, or humans.
Tool sandbox: many agent calls just need a safe place to run a tool call (run a script, check code, transform files) and return results — no long-running home needed. Claude Managed Agents can run its tool calls in sandboxes on your own infrastructure, which kagent supports.

Key 1.0 designs (see image): Sessions live in a database naming one Actor; model keys and MCP headers are injected on egress so the runtime sees placeholders; Agent + Harness map to ActorTemplate with conversations pinned to revisions; Actors suspend while waiting on users and reading history never wakes them; an A2A gateway decides who may talk to which conversation before handing off to Substrate's router.

Compared with running agents as Pod replicas, 1.0 adds isolated sandboxes, efficient scheduling, persistent agent state, and production-ready security.

Original post →

kagent 1.0 GA: Agent Substrate brings sandboxed AI agent execution to Kubernetes beyond pod-per-agent

More from coding & agent