How Pushary for Teams decides who an agent asks

Pushary for Teams sends each AI agent's permission question to the person running that agent. A rule can send a step you choose, such as a push to your infrastructure repository, to a lead instead. Answers given in the Pushary phone app, Mac app or dashboard are saved with the person's name, so the team can see whose agent asked and who said yes.

Why "who gets asked" changes on a team

On your own it's easy. Your agent stops, your phone buzzes, you decide.

Put five people on one workspace and it isn't. On Pushary's Agent Pro plan, every member gets every question. For two people at the same desk that's fine. For five, it means four people get a buzz for a git push that isn't theirs, and nobody is quite sure who should answer. So the question sits. Meanwhile the agent that asked has stopped, and the person who started it is in a meeting.

Then something on main looks wrong a day later, and the team chat asks the obvious thing. Whose agent was it, and who said yes?

The agent asks the person running it

Once a second person joins the workspace, each question goes to the person whose agent asked, on their phone and in their dashboard. Tom's Codex session asks Tom. Ada, an admin on the team, runs Claude Code, and it asks Ada.

Pushary knows whose agent is asking because each person signs in and connects their own agents, on their own computer. Before the second person joins, questions also go to browser push, and to Slack if you connected it.

Task updates work differently. A task update is a message the agent sends on its own, such as a note that a long build finished. Those go to everyone's phone, not only to the person whose agent sent it. The Team docs give examples.

A step you choose goes to a lead

Some steps shouldn't be decided by whoever happens to be running the agent. A push to the repository that holds your infrastructure is the usual one.

An owner or admin opens Policies, finds Approval routing and adds a rule with three parts:

• The step, as a tool pattern, for example Bash(git push:*) for every git push.
• The repository, one or all of them.
• Who is asked, one person or a role such as Admins.

With a rule for git push in infra pointed at the Admin role, Tom's agent asks everyone with that role (here, Ada) before it pushes there, not Tom. The owner is a separate role, so pick Owner too if they should be asked. Pushes to other repositories still ask Tom.

Only they can answer is on by default. Pushary asks Ada first and refuses an answer from anyone else. If Ada hasn't answered when the wait ends, the step goes back to Tom's agent. Where the agent has its own prompt, Tom can say yes there, and Pushary doesn't record that answer. A rule decides who is asked through Pushary. It doesn't lock Tom's terminal.

Rules match tool approvals only. If the agent writes a question of its own, such as which of two migrations to keep, it goes to the person whose agent asked.

An empty rule steps aside. If nobody holds the role any more, or the person it names has left, the rule behaves as if it weren't there and the question goes to the person whose agent asked, so no agent waits on a role nobody holds.

One limit to know before you rely on a phone. The phone app can approve a plain push. It can only deny risky steps such as a force push, rm -r or a deploy. Those are approved in the dashboard, or at the agent's own prompt, where the yes isn't recorded.

One record of whose agent asked and who said yes

Activity, in the dashboard sidebar, has three tabs: Sessions, Audit trail and What changed.

Sessions shows one row for each session your team's agents ran: whose agent it was, and the commands, file edits, commits and questions it reported. Read-only commands are counted, not listed. Worth a look sits on top and picks out items such as a very fast yes on a risky step. You can narrow it to this sprint or the last 7 or 30 days, and to everyone, yourself or one person.

The ⋯ menu has Copy for retro, Export CSV and Export JSON. Owners and admins also get the sprint record by email on the sprint's last day, except during the trial. History is kept for 365 days.

Everyone in the workspace sees the same pages, and nobody is scored or ranked. The point is that a retro can start from what the agents reported and who answered, instead of from what people remember.

For the wider argument about why that record matters, see who is accountable when an AI agent makes a mistake. For the fields and exports, see the audit trail docs.

What the record does not show

• Who said yes at the agent's own prompt. Activity can show that a step was left to the terminal, and the action if the agent reports it, but not who said yes.
• Who answered outside the Pushary apps. Answers in the phone app, Mac app or dashboard carry a name. An answer from a link in a browser notification or the web inbox doesn't.
• An answer for steps your rules approve on their own. They run without a question, so there's no answer to show. The actions they report still appear.
• Full commands. A command shows its first two or three words, such as git push.
• Anything from before a person connected. Each person's record starts when their agent first connects.

So if the record matters for a step, answer it in the phone app, the Mac app or the dashboard. The Team docs keep this list current.

Where Teams fits

On Agent and Agent Pro, the agent asks you. On Team, it asks the person running it, or the lead a rule names. With Partner APIs, an agent inside your product asks your own customer. The manifesto explains why the person running an agent isn't always the one who should decide.

Set it up

1. Start the trial on the Team page. New workspaces get 3 days free.
2. On the computer where your agents run, run npx -y pushary@latest setup and scan the QR code with the Pushary phone app. On Windows, Overview shows the PowerShell command. On a Mac you can use the Mac app instead.
3. Invite your team by email from People on Overview, the dashboard's home page. Each person then connects their own computer and phone.

Routing rules are optional. Add one when there's a step you want a lead to see first. Each person still picks their own approval mode, and the Team docs cover roles, seats and billing.

Team works with Claude Code, Codex, Cursor, Gemini CLI, VS Code, OpenCode and Hermes. It costs $59 a month for 5 people, plus $12 a month for each extra person.