# The coworker and AutoMode

Volae has one agent, and it's the same one everywhere: Today, the side panel, ⌘K and Pro lookups. It stays on and works toward your goals. You handle what needs judgment.

## How it works

- **Events wake it.** A new message, a negative spike, a failed payment, a ranking drop, a rival's price change, a theme marked shipped. Rules and a small model triage each event; only what matters wakes the full agent. Wake-ups and spend have a daily limit; anything over it waits for the daily review.
- **The daily review.** Once a day, in your time zone, it catches up, runs GEO and SEO checks, looks at goal progress and plans the week.
- **Goals.** Give it something to win ("rank in ChatGPT for this question", "raise trial conversion") with `set_goal`. It breaks the goal into work across days and keeps notes on what it tried.
- **Subagents.** For big jobs it splits the work and runs it in parallel. You see progress and results, and there's nothing for you to manage.
- **Say what you want to see.** It shows tables, charts, lists of people, posts and comparisons on the spot, built from a fixed set of components and checked data. It never runs code the model wrote. Pin a view to Today or save it. Changing a view needs no approval; changing data or sending anything goes through the normal rules.

## Today is a standup

`get_standup` is where to start each day. `get_today` is narrower: the daily run's own report, its step-by-step status and when the next run starts. `get_standup` and the Today page show four things: **what I did** (each item can be undone, and auto-approved ones are marked), **what needs you** (only what really needs a decision, with a suggestion), **what I'm working on** (each goal's progress) and **what I plan this week** (you can veto any of it with `veto_plan_item`).

## AutoMode

Each action the coworker wants to run is either cleared automatically or handed to you. It goes through these checks in order, and stops at the first one that decides:

1. **Hard limits.** Destructive actions, internal changes that can't be undone, spending over a threshold, and first contact with a stranger always need you.
2. **Always a person.** Social posts and public replies, anything sent to more than one person, follow-ups and outreach, publishing to your help center or opening a pull request, and knowledge base updates always need you, however confident the system is.
3. **Outside content.** An outward-facing action whose trigger touched outside content (an email, a web page, a post) is never automatic. Setting that action to automatic doesn't change this.
4. **Your settings.** If you set an action to "off" or "ask me", that's what happens. "Automatic" only applies after the checks above.
5. **The judgment model.** For the rest, a small decision model answers a fixed set of questions: is this within your goals and instructions, is it the same shape as what you've approved before, is the damage small if it's wrong? It only clears an action when all answers are confident. It never sees the original text of emails, pages or posts, only your goals and instructions, the action's metadata and its history. If it fails or times out, the action goes to you.

Everything that runs goes in the change ledger; what can be undone has an undo button. Undoing or editing what AutoMode cleared makes that kind of action less likely to clear next time. `list_auto_decisions` and `get_auto_decision` show why each decision went the way it did.

You can turn AutoMode off in **Settings → Teammate permissions** and go back to the plain table of automatic, ask me and off per action.

## Budget

Give it a monthly credit budget in **Settings → Teammate permissions**. It reports what it spent and what that bought. A task over its budget, or any one task over 50 credits, gets a quote first.

Included in your plan: event triage and the daily review. Charged at actual usage: waking the full agent for an event, chatting with it, showing or changing views, and the tasks it does. See [Credits and billing](credits.md).
