Parallel agent tasks in git branches

When a team has five developers and each one has an AI agent, five tasks are running in the project at once, each in its own branch. Their code is kept apart, and git has done that well for years. But the project knowledge the agents get is usually shared.

The problem

Take two tasks. In the first, the team decided to move payment retries to a queue. In the second, a developer is fixing a report that reads the same data.

If the decision from the first task goes straight into shared memory, the agent in the second task starts relying on it. But the first branch hasn't passed review yet. The decision may be revisited or reverted, and the second task is already written around it.

The reverse happens too. The agent in the second branch doesn't know about a decision that has already been made and merged, because its memory wasn't updated.

How we solve it

In AI Council, a task is a git branch. Knowledge that appears in a task is tied to its branch:

• Until the task is merged, only the agent working in that branch sees its decisions.
• After merge, the knowledge moves to the target branch, and everyone working with that branch sees it.
• If the branch is closed without a merge, its knowledge never reaches the shared base.

Team rules are shared across the whole project: only knowledge lives on branches.

The same principle applies to review. AI Council computes knowledge and checks from your working tree, not from a branch someone typed in by hand. If new commits land on the branch, the review is marked outdated.

What your team gets

Agents on parallel tasks don't get in each other's way. Decisions go through the same path as code: a branch, review, merge. And the knowledge base doesn't fill up with ideas that were eventually dropped.

How AI Council works