This request matches a reliability problem we have been working through in a multi-agent setup outside Dots.
Shared context is useful, but in practice it is not enough by itself. The parts that made handoffs reliable for us were:
- explicit task ownership, so exactly one agent owns the next state transition;
- opaque thread/task IDs instead of agents inventing or reusing identifiers from prose;
- idempotent request keys, so retries do not silently create duplicate work;
- selected shared context rather than an undifferentiated common memory;
- capability/permission checks at the executing agent, not assumptions carried through the handoff;
- no automatic propagation of operator authority just because another agent says “the user approved this”.
That last point matters a lot: shared context and shared authority are different things.
I would also strongly support a capability check before a dot presents a team design as executable. Planning “five specialist dots with assistants” is fine, but the UI/runtime should clearly distinguish “proposed organization” from “objects this account can actually create”.
A shared task board with owner, state, handoff lineage, permissions and retry/idempotency state would solve a large part of the coordination problem described here.