Skip to content

The Agent

The Agent is a program that reads and changes a Project on a Member’s behalf. You ask it in words; it changes scenes with the editor’s own tools, and writes the game’s code.

The Agent panel is on the right of the editor, beside the Inspector. It holds everything that was asked of the Agent in this Project, in order, and every Member sees the same list.

The Agent is off until an Admin of this host gives you an allowance. Until then the editor has no Agent panel, in your browser and in the Shell app, and Agent on the Dashboard says Off; you can run an agent on your computer instead, at any time. A teammate’s Run still shows as Agent among the Members in the toolbar, and its changes appear in the scene.

  1. Type what you want in Ask the Agent…
  2. Press Enter, or click Ask. Shift+Enter makes a new line.

Your prompt appears in the Agent panel with your name. The Agent’s work on one prompt, from the prompt to its end, is a Run.

A prompt can point the Agent at an entity, a script or an asset. The Agent is told exactly which one, by its uid or its path.

  • Drag an entity from the Hierarchy, an asset from Assets or a script from Scripts onto the Agent panel. It shows what it would point at before you drop it.
  • Type @, or click the @ button, and pick from the menu: entities, prefabs and scenes, scripts, pictures and sounds. ↑ and ↓ choose, Enter adds, Esc closes.
  • Selected: above the prompt box adds what you have selected in one click.

Each thing you point at is a chip above the prompt; its × takes it out. In the Agent panel, the sent prompt shows them as chips in its words. Click one to select the entity, open the script, or open the asset. If it is gone since, the panel says so.

It saysMeaning
Waiting for its turnAnother Run is under way. One Run is done at a time, in the order asked
WorkingThe Agent is at it. The bar above the prompt box says so
DoneThe Run ended, and the entry closes with what the Agent says of it
FailedThe Run ended with a problem, which the entry names
StoppedSomeone stopped it

While a Run works, a bar above the prompt box says Working, how long it has worked, the model, how many prompts wait, and Stop. Every Member sees the same time, also one who opened the Project after the Run began.

An entry shows what the Agent says as it goes. The steps it takes between are folded into one line that says what they did, such as 8 steps · Read the Gamedoc, added @shell/particles: click it to see them. Each step says what the Agent did and to what:

  • Read the Gamedoc, Looked at the editor, Added @shell/particles.
  • Wrote bat.system.ts +48: the lines it added, and taken away (−2) where it edited. Open in Code opens the script in the Code panel.
  • Checked the code: 2 errors, with the errors under it. A step that failed says failed, and why.
  • Steps of one kind in a row are one line, with what they made and the time they took: Made 7 sounds 18s. A step taken again in a row shows once, with a count (×2).

The step under way has a spinner. Once a Run has ended, what it did before its last words is one line, such as Worked 4m 38s · 31 steps: click it to see it all.

Each entry before the newest is folded to one line: what was asked, how it ended, and how long it took (Done · 1m 04s). Click it to open it.

During a Run the Agent is in the editor like anyone else: it shows among the Members in the toolbar as Agent, and what it changes in a scene appears as it changes it.

When a Run changed the game’s code, your page loads anew once, after the Run has ended.

Stop, in the bar, asks the Run under way to stop. The bar says Stopping, then the entry says Stopped. What the Agent had done by then stays.

A prompt is asked with what the Runs before it said, so “make it faster” is understood from what came before.

Tick Begin anew to ask without that: for something unrelated, or when the Agent holds on to an idea you have left. The tick is for one prompt and clears itself.

The dropdown beside Ask chooses the model the next Run is on: Opus 5.5, Sonnet 5.5, or Haiku 4.5. Opus 5.5 is chosen until you choose another, and your browser remembers what you chose.

Each Run names the model it was on.

A Run that has ended shows what it cost:

Done, Opus 5.5, $0.79

Under the prompt box is what you have spent this month of the allowance an Admin gave you, and the day it is counted again from, as your last Run said it. It links to Usage on the Dashboard’s Agent page:

$12.40 of $25.00 until 1 Nov
  • The amounts are in US dollars. Less than a cent shows as “less than $0.01”.
  • Allowances are per Member, not per Project. Your Runs in every Project count toward yours.
  • The month is the calendar month.

When your allowance is spent, the panel stays, a Run fails, and its entry says from which day you can ask again. Everything else in the editor works as before. An Admin can raise the allowance.

With no allowance, a Run that is asked all the same fails with You have no allowance, so the Agent is off for you. Ask an Admin of this host to give you one, or run an agent on your own computer: see Agent on your Dashboard. When an Admin takes your allowance back, a Run that is working fails at its next step and says why, and the Agent panel is gone when you come back to the editor, or open a Project. When an Admin gives you one, it comes the same way.

A Run is stopped once it has cost about $10, whatever the Project’s files say: it stops where it is and fails with Reached maximum budget ($10). Ask again to carry on: the next Run has $10 of its own, within your allowance.

Where a Project has a Gamedoc, the Agent reads it before it builds: it says what the game is to be. Where what you ask in the Agent panel says otherwise, what you ask counts. Where what it builds changes what the Gamedoc says of the game, it changes that part, and says so at the end of the Run. What it changed is marked in the Gamedoc.

The Project is kept as it was before every Run: a Checkpoint. See Checkpoints and Versions.

The Agent works inside the Project. It cannot invite anyone, and it reaches nothing on the Dashboard.