Upgrade without losing control

Treat an upgrade as organizational change

A running Fleet has several version states:

writable AgentOS checkout
loaded harness instructions
database migration journal
running image digest

Changing one does not change the others. A Git checkout can be current while a native session still has older instructions. New Markdown can be present while executable runtime code remains in the old image.

First Mate may discover a newer stable release read-only. It should report the exact revision and relevant changes without modifying a dirty checkout or restarting itself. After approval, the released procedure verifies provenance, inspects intervening migrations, preserves PVC and recovery boundaries, advances ordered state and verifies each Mate against the intended version.

An upgrade is complete when the same organizational identities, homes, work and authority remain intact under the new release—not merely when a new image becomes Ready.

Preserve sovereignty through native authorities

Sovereignty means people control the technology they create and depend on. AgentOS keeps durable company state in standard systems the operator can inspect, self-host and replace:

• Kubernetes for workloads, authorization and storage;
• PostgreSQL for organizational coordination;
• Herdr for native sessions;
• PVCs for unfinished homes and worktrees;
• Git and providers for delivered source.

A user interface, protocol, event bus or observability warehouse can improve how humans and Agents collaborate. It must not become the only place the company knows itself or authorizes effects.

The support company might replace Discord with a self-hosted, Buzz-like interface and connect additional harnesses through ACP. That migration should not rewrite Task identity, Agent homes or Git delivery. The new surface submits intent to the same First-Mate boundary, reads approved projections and routes exact decisions. If it disappears, the native authorities still preserve the organization.

The autonomous company you now know how to build

You began with one prompt and one persistent First Mate. You delivered a bounded outcome, protected human attention, recovered work, added one durable domain, connected external reality and built a reviewed learning loop.

That is the AgentOS path: independent systems that follow human intent and stay within boundaries people set. Return to the course introduction, use the technical documentation, or bring AgentOS online.