Agents read, people decide

An AI agent connected to unbranch is a capable collaborator. It reads the whole model, the ledger and every proposal in review. It can turn a half-formed idea into a well-structured draft. It still cannot make the decision.

What an agent can do

• Read the model, the proposals, the commitments and the ledger of every project you can open.
• Talk an idea through with you and say where it belongs: already there, a change to a capability that exists, a new capability, a new direction, or not a product change at all.
• Write a draft with you, check it, and send it for review when you tell it to.
• Discard a draft it wrote.
• When you choose to change a proposal already in review, write its next version and leave it waiting for you. The proposal itself does not change until you submit that version on the web.

What an agent cannot do

• Consent, abstain or object.
• Accept or reject a proposal.
• Start delivery or verify it.
• Change or submit a proposal that is already in review.
• See or touch anybody else’s draft.
• Reach a project you cannot open.

No tool exists for any of these, and none is planned. They are decisions about your product, and decisions belong to the people who answer for them.

Why it is built this way

• Accountability needs a person. Consent means “I have no reasoned objection”, and only someone who will live with the result can say that.
• Attribution stays honest. A draft an agent wrote is marked as written by an agent. When it sends a draft for review, the ledger entry records the actor as an agent. Your account owns the action, but the record shows the words were the model’s.
• The record stays trustworthy. The model is worth reading only if everything in it was agreed by people. One automated acceptance would make every item suspect.

What this means for you

Use the agent for what it is good at: reading everything quickly, finding where an idea fits, and drafting carefully. Keep the decisions for the team. See Working with an agent.