Agent systems need a memory, and a boundary.
Durable handoffs, explicit authority and evidence of completion are part of the engineering.
A conversation is not a work ledger.
Useful work often crosses sessions, tools and machines. A result posted in a conversation can still be lost to the wider project. A handoff that lasts needs someone to receive it, a link to where the work is, and a clear record of what happens next.
That question drives Wild Ducks’ own agent research: how can agent activity stay connected to the commitments and decisions of the person it works for?
Make the authority concrete.
An agent needs to know what it is allowed to do, what resources it may use and when that authority expires. Permission should travel with the work, rather than be inferred repeatedly from conversational context.
Read-only research, reversible implementation and an external commitment have different consequences. The system should represent those distinctions and provide a way to pause, cancel or request a decision.
Completion needs evidence.
A claim that something is done is weaker than a result you can open and a check that it is right. Keep references to the source records and artefacts, keep track of where they came from, and tell an attempted operation apart from a confirmed outcome.
These are design principles, not a promise of perfect autonomy. Independent evaluation still has to establish where an agent system is useful, where it fails and whether additional coordination earns its cost.