Updated System Instructions to avoid bad state/conflict #17753
Description
The primary goal of this issue is to add some level of steering for the main agent to help steer the model from initializing subagents that will most likely have conflicts and cause a bad state.
This is not the end state solution but an attempt to rely on the model's reasoning to avoid obvious bad state. In this early version of parallel agents, users must also be thoughtful in how these agents are invoked. We may have more complex locking/worktree/etc solutions to solve this.
Criteria:
- Update system instructions to avoid obvious bad behavior
- Add appropriate behavioral evals to verify that the additional (or not needed) SI changes help the model avoid bad subagent invocations (ex. use the same doc string agent to modify the same file).