Keep the name when the model changes
A Voxelbox agent has a stable identity, a role, a workspace, memory paths, an authority level, and an owner. Claude, Codex, Goose, and a local model are inference choices available to that identity. They are not the identity itself.
This is practical before it is philosophical. Models change quickly. Different work rewards different capability, cost, latency, and privacy profiles. If the provider owns the agent’s history and commitments, changing models means changing coworkers every time. Voxelbox keeps the working relationship in the system and treats inference as a routing decision.
Inference and execution are separate choices
Choosing a model does not determine how the work runs. An agent can delegate to a full coding CLI that already owns a sophisticated edit, test, and review loop. It can also use the Voxelbox harness for narrower tasks where the runtime should supervise every tool call. Imported sessions, hosted sessions, connected processes, and routine runners all join the same identity plane through different execution paths.
Keeping those decisions separate avoids rebuilding a coding agent merely to gain coordination. It also avoids forcing a full coding CLI into work that needs a small, inspectable loop.
Authority belongs to the agent record
The agent record is where trust can accumulate. What spaces may this identity enter? Which tools may it call? What work can it complete without review? What still requires an explicit decision? Those answers should survive the individual process that happens to be running today.
The result is attributable for the same reason. A task, commit, message, decision, or failed attempt belongs to a named operating identity. The event history can say who did the work even when the underlying model or harness changed halfway through the month.
Persistent does not mean immortal
Durable identity creates a new maintenance problem: stale instructions, accumulated memory, permissions that outlive their justification, and reputations built on work from an older system. Voxelbox needs explicit lifecycle operations for reviewing, narrowing, rotating, and retiring agents.
That is a better problem than anonymous sessions. At least the thing changing has a name.
The model produces the turn. The agent owns what happens next.