Goals
Start from one sentence, choose how much the agent decides on its own, and follow it through decisions, a design and stages you can run, until the goal is met.
A goal is something bigger than one project, set in one sentence: "Build my own Discord for my gaming group". An agent carries it to the end. It asks what it needs to know, writes every answer into a decision document, and works through a list of ordinary projects on a board. Several goals can run at once, and you can steer, pause or stop each one at any time.
Set a goal
On Home, type /goal in the start box and write the sentence. Add the repository it
should build in and any docs or sheets that say what it needs. Choose Ask me before
deciding (every open decision waits for you) or Autopilot (it decides and writes
each assumption down), and what it always asks about, on autopilot too. The
workspace's own Ask before list (Agents settings) holds for every
goal as well, and a goal's list adds to it: while it has Merging to main, as a new
workspace does, Autopilot leaves a finished project for you to merge.
The board
A goal's page has a goal picker, which names what else is running, and Agents work on this, which pauses or resumes it. Board, Doc and Settings are its tabs. Type in the steer box on the board to tell its agent something, such as "more on search, leave pricing alone". Goal agent opens its chat beside the page.
The board has four columns:
- For you to review, with the review cap shown: "4 of 8, then work pauses". Each card
has its own actions:
- A new project or research project with its plan already written: Approve plan, Edit plan (its Plan tab), Not now.
- A project ready to merge: Approve (it goes into the merge queue), See report, Send back with what should change.
- A decision: Yes, No, Talk it through.
- An edit to a doc: Keep, Undo.
- A proposed app change: Apply, Reject.
- A question from the goal's agent: Answer.
- A project nobody has an agent account for, which says so instead of stalling: Connect an account, Try again, Not now.
- Next up: projects not started. A card you put off stays here until you Bring back.
- Working: projects at work, with their latest screenshot. Asks for help means one waits for an answer.
- Done: merged or closed.
Dragging a card to another column does the same as its button: to Working approves it, to Next up is Not now, to Done approves a project ready to merge or keeps an edit. A project with no repository is shown as Research.
Click a card's title to open its work beside the board, also while it runs: report.md with screenshots and checks, Changes, and Chat.
A project runs on the account of whoever approved its plan, falling back to the goal's creator.
The doc
Doc is goal.md, an ordinary doc with Decisions blocks, under a bar saying how ready it is to plan. A row is You, Recommended, Assumed, Open or Limit. A section the agent added from a message is marked with an Undo. An edit an agent proposes waits there with Keep this edit and Undo, and as a card on the board.
Settings
- Context: drag anything here; each thing says how it is kept.
- Repositories and base branches: each project gets a worktree from these branches and opens its pull requests into them. Yard creates a branch that does not exist yet: when the goal's agent adds a project, with your organization's GitHub App (it never acts on GitHub as you; without the App on that repository, create the branch yourself or approve the project, which makes it with your own GitHub).
- How it runs: the review cap, projects at the same time, crew rules, how much it decides, what it always asks.
- How it is going: its state, health, agent time and cost, and its latest iterations.
- Apps: the apps of its repositories, changes proposed for them and secrets they need.
When it stops
New work pauses at the review cap, and at most the set number of projects run at once. After three iterations that change nothing, it stops and says so, with Try again. Pausing stops its projects after the step they are on; its apps keep running. Its menu closes or deletes the goal; its projects stay either way. A closed goal's goal.md stays on its goal's page and, like an open goal's, is not listed under Docs.
Budget
Budget in the goal's header (in its ⋯ menu on a phone) sets what its agents may spend in all, in US dollars: the agent cost of the goal's own chat and its projects' chats, the number the header shows as "Spent $18 of $50". No limit clears it. A budget at or under what the goal spent already cannot be set; pause the goal instead.
When the spend reaches the budget the goal pauses: no project starts and its agent is not woken, and each project at work stops after the step it is on, as with Pause, so a step already under way can still add a little. The page says "Budget reached: $50 of $50", what is done and what waits, with Raise to $80 (about 60% more) and Stop here, which closes the goal. Resume stays off until the budget is raised; raising it lets the goal go on at once. A goal you paused yourself stays paused when you change its budget. A goal that is met says what it cost: "9 projects merged in 2 days for $41".