Assigning work is not starting an agent.
Responsibility can be durable before execution is safe, ready, or even possible.
Planning and release are different acts
A founder may decide that Operations owns an investigation while the Task is still missing logs, blocked by another change, or waiting for a maintenance window. The assignment is already useful: it establishes accountable ownership and lets related work be planned. It should not silently launch a process.
Execution requires another set of conditions. The Task must be in a ready state, blockers and waiting conditions must be clear, the relevant Project must permit the work, and a compatible worker must be available. Keeping those conditions explicit makes automation understandable instead of surprising.
An Agent Actor is not a process
An Agent Actor such as Development, Research, or Operations is a durable accountable identity. It can receive Tasks, have scoped instructions, leave attributable work, and build a history across many concrete sessions.
A worker and harness are execution machinery. They may appear and disappear, run on a laptop or VM, and use different models over time. Confusing the Actor with the process makes authorship unstable and turns every infrastructure change into a change in who supposedly owns the work.
Assignment says who owns the work. Readiness says whether an attempt may begin.
The separation improves failure handling
If a worker disconnects, the Task does not lose its owner and the business does not need to invent a replacement task. Ro can retain the assignment, mark the run as lost or uncertain, preserve the last report, and ask a human whether to reconnect, wait, cancel, or start a replacement attempt.
It also prevents a dangerous shortcut: automatically redispatching work simply because a lease expired. The original external process may still be running. Repetition should happen only when policy established in advance says the operation is safe and idempotent.
Review remains part of the work
Finishing an agent run does not finish the Task. The result may need a named reviewer, an explicit decision, or a material edit that invalidates an earlier approval. Run state and Task lifecycle are related, but they are not the same axis.
This distinction sounds small. In practice it is the foundation for accountable human and agent collaboration: durable responsibility, controlled release, inspectable execution, and review all describe different truths without forcing one status label to pretend they happened together.