Skip to content

Boundaries

Last updated: 2026-07-30

microagent runs Linux workspaces inside microVMs, and it stops at the VM boundary. If you are building a runtime on top of it, this is the line to keep in mind.

microagent supplies the VM layer: kernel, rootfs conversion, lifecycle, state, and structured CLI/MCP output. Your program supplies identity, policy, credential authority, and intent.

  • VM commands (run, create, start, status, halt, stop (CLI aliases it to halt), quarantine, pause, resume, kill, delete)
  • OCI image to ext4 rootfs builds
  • Identity in requests and state files
  • State changes as JSON
  • Readiness, structured exec, structured results, and declared artifacts
  • Host supervisor boundary
  • MCP stdio adapter over the workspace APIs
  • State files and cleanup
  • Host/guest wiring such as vsock listeners
  • Planning loops
  • LLM/provider calls
  • Tool mediation and tool policy
  • Policy decisions
  • Audit meaning and retention
  • Credential eligibility, authorization, and grants
  • Agent frameworks and user experience

Identity, policy, and credentials stay outside

Section titled “Identity, policy, and credentials stay outside”

microagent transports identity; it never mints or judges it. Every request carries an identity block that is recorded in state files and events (see State and identity). But the meaning of a role, the decision to allow an action, and the authority behind it belong to your control plane.

Tool mediation follows the same rule. The mediation channel gives the guest one declared path to your host control plane; your listener decides what each call may do. For credentials, microagent implements mechanisms: it can resolve an operator-declared reference, deliver a secret, or mechanically substitute a host-held credential into a request. It does not decide whether the caller is eligible to use that credential or what the resulting access means. Your secret manager remains the source of truth, and your control plane owns those decisions.

If a guide asks you to write a policy check, a host listener, or a credential authorization decision, that belongs outside microagent. It is not a missing feature.

  • Public output is structured and machine-readable.
  • AX (agent experience) responses over MCP are typed and decision-relevant instead of requiring CLI log scraping.
  • State changes are API output, not log strings.
  • Identity is preserved explicitly in requests, state files, and events.
  • Host details stay behind supervisor boundaries.
  • Invalid VM config fails closed.
  • Narrow protocols beat shell-string execution.