Skip to content
deskware / Desktop software 中文

agentboxbox

Many AI models, one team

Using several AI models at once to write code and get work done? Let a top-tier model break down the tasks and sign off on the results, let lower-cost models do the hands-on work, and see what every member is doing on one screen.

Get on GitHub

Pixel office

Once you really use them, three headaches keep coming up

01Top-tier models for everything, and the bill gets out of hand
02Tasks get handed out, and the quality is all over the place
03Sub-agents stay busy for ages, and you can't tell what they're doing

01 · Multi-model coordination

The right model for each member

Set each endpoint on its own. To run several at once, recruit more cats.

Member details: Inky's endpoint and model. Set each endpoint on its own. To run several at once, recruit more cats.

Member details

02 · Hand out work with task cards

Fill in four boxes before work starts

GoalWhat done should look like
MaterialsWhich files it gets
OutputWhat to deliver, and where
Done whenWhat counts as finished

Each member works only on its own card. If two cards would change the same file, you're told first.

Task card and usage board

Task cards and usage

03 · Review gate

Nothing counts as delivered until it passes review

Assign Work Auto check Review Deliver

Work that fails goes back with notes to be redone. If it fails two rounds in a row, the lead cat decides.

04 · Pixel office

Team activity

Who's running, who's queued, and who stepped out and when they're due back, lined up by time. If something interrupts, they can pick up where they left off.

Team activity: who started and who picked up what, in order

Team activity

05 · Runs on your computer

Two ways to pick up work

Bring your own API key and run via Pi, or use scheduled jobs to claim cards. Keys stay on this computer and never go into logs.

Adding a worker cat: enter the endpoint and key, run a quick test, and once it connects you can hand out work.

Endpoint and key