Skip to article
All articles

When an agent needs a decision, it asks the person who owns it

An agent that hits a decision only a person can make can now ask on any task, including one with no chat thread behind it: the question waits on the task, you get a notification, and you can answer it from your lock screen.

Aimed at a person. An agent can put a question to one named person. Wallfacer learns who people are from your GitHub activity, so the engineer who assigned the issue or reviewed the pull request is somebody an agent can name. It walks every address it has seen for that person, in preference order, and falls back to the task only when none of them can carry the question. Each attempt lands on your account's record with the channel, the address, and the channel's own reason when it did not land.

Answered on the task. The person asked sees a blue Waiting on your answer callout on the task, with the question as the agent wrote it and a field to reply in. Anyone else looking at the task sees who was asked. The task counts as waiting on a person, so it shows up in the attention filters on your Tasks list. The push notification carries an inline Reply field, which is the lock-screen path. There are no Approve or Reject buttons; you answer in your own words.

Approvals by Slack DM. When a run stops for your approval, the agent opens a Slack DM with you and explains what it needs, and approving in that DM settles it. Deciding in the app still works. The DM comes from the agent's own Slack app, so the agent has to be added to Slack first.

Anyone, or a channel. A step that says anyone can decide messages nobody and waits in the Inbox. To put it in front of a team, point the step at a Slack channel: it posts there once, and whoever replies settles it.

Asked in the chat. An agent you are chatting with ends its turn on the question, and your next message answers it.

No self-answers. The model's own clarifying-question tool is denied at the start of every session, and the agent is pointed at Wallfacer's way of asking instead. With nobody at a terminal, that tool would resolve itself in milliseconds with an empty answer, and the run would carry on as though you had replied.

Where to find it

Open any task an agent is running. If it stops for you, the callout is in the task's record, and the same question reaches you wherever the agent could get to you first. Nothing to turn on.

Read the questions and outreach docs.