The agent should not be the durable unit of work. The work should be.

We keep focusing on making agents smarter and more autonomous. Frankly the more I have worked on building midfleet the less I worry about what model the agent is using.

I think the harder problem was making the work survive the agent. I.e. I needed the mandate I gave the agent, to survive regardless of what happens to it.

Agents will stop. They’ll hit limits, crash, lose context, or get replaced by a better model.
Completely normal.

The goal, owner, boundaries, decisions, progress and proof of completion should still be there.

Then one agent can implement.
Anpther one can review.
A human can approve.
Another agent can finish.

The worker can change without the job becoming an orphaned chat.

Agents should be replaceable. Responsibility shouldn’t be.