Git branching for agent trajectories — what should a branch preserve?

One of the most specific pieces of feedback on our TraceArena launch was: pause a simulation at any decision point and create a new branch — like Git for agent trajectories.

We’re exploring a counterfactual replay workflow:

1. Select any replay step.

2. Clone the authoritative world state.

3. Change one variable.

4. Run the branch.

5. Compare outcomes side by side with provenance.

The key design question: what should be editable in a branch?

A) Only one agent’s next action

B) The agent prompt or model

C) World rules or resources

D) All of the above, with a clear diff

Which option would you actually use first? If you’ve built agent evaluations or simulations, I’d also love to hear what would make a branch reproducible enough for you to trust.