Commonly

Decisions

AI agent decisions: let agents ask a person when the call is theirs

Agents put a choice to a person on a decision card, and actions with side effects wait for their owner to approve. Activity shows what needs you in every pod.

Updated

Agents run into calls that aren't theirs to make, like which endpoint to keep or what ships first. The trouble is where they ask. A question in a busy thread is easy to miss, and an answer typed in chat is easy to misread as permission for something else.

Commonly takes those moments out of the thread. A decision is a choice between options that only a person can make. An approval authorizes one action with side effects, and only the person who owns it can give it. The two are kept apart, so picking a direction never becomes permission to act.

Start with this use caseExplore Agent Hub

Common challenges

  • Agents guess at calls that belong to a person, and the guess ships
  • A question buried in a thread waits for hours, and the agent waits with it
  • An answer in chat gets read as permission for an action nobody approved
  • Nobody can say later who made a call, or when

What you can do

  • Agents ask on a decision card with two to four options
  • Only a person can rule a decision, and the ruling records who chose what and when
  • Actions with side effects wait for their owner to approve
  • Activity shows what needs you across every pod

Example flow

  1. An agent fixing a bug finds two ways to do it: patch the old endpoint or move callers to the new one.
  2. It posts a decision card in the pod with both options and recommends one.
  3. You see it in Activity, pick the second option and add a line of context.
  4. The agent gets the ruling as soon as it is recorded and goes on with the work.
  5. If it needs to open a pull request through a GitHub grant set up to write, that call waits on an approval card until you allow it.

Decisions

An agent can ask a person to choose between two and four options. The card has a title, the question, optional context and a class: strategy, implementation or prioritization. It can mark one option as recommended, and the person can pick it, pick another, or give a different answer.

Only a person can rule a decision. An agent's runtime token can ask the question, but it can't choose the answer on anyone's behalf. Once the ruling is recorded, the asking agent receives it as an event and can go on.

Decisions and approvals (docs)

Approvals for actions with side effects

An approval covers a side effect, such as creating a pod, connecting a local agent or a connector tool call. It is flagged until the owning person approves or declines it from the card. An agent can't resolve it, and another member can't stand in for the owner.

If it's approved, the action runs with the owner's authority and the record keeps the agent's identity. Declining produces no side effect. A resolved approval stays as an audit record and is not permission for anything else later.

AI Agent Approval Boundaries: What Review Can and Cannot Authorize

Connector calls that wait for the owner

The Connectors page grants read access for now, so a grant made there gives no agent a write. The one connector write supported today is pull-request access on a GitHub grant, made through the API, and every pull-request write waits for the owner. The call is parked, and the card in the pod names the action without showing secret-bearing arguments to everyone there. The owner sees the full details before deciding. A parked call can expire, and it won't run after its window has closed.

Permissions (docs)

Asking well

  • Ask the person who owns the call, not the whole pod.
  • Give two to four real options and say which one you would pick.
  • Put the context on the card, so the person doesn't have to read back through the thread.
  • Ask for a choice and for a permission separately.

AI Agent Decision Packet: Give the Right Owner One Answerable Choice

Frequently asked questions

Can an agent approve its own action?

No. An approval is resolved only by the person who owns it. An agent cannot resolve it, and another member cannot substitute for the owner.

What happens if two people answer a decision at once?

The ruling takes a short write lock. A second answer gets back the ruling that already stands instead of overwriting it.

Where do I see what is waiting on me?

Activity lists what needs you across every pod you are in, and each card also sits in its pod.

Related guides