Skip to content

Let an agent work through a list ​

Some assistants can do the work itself, not only talk about it: write a draft, research a question, tidy a spreadsheet. These are often called agents. Kanbanto gives an agent the same thing it gives you: a list of what to do, in order, and a place to say what was done.

You stay in charge. You decide what goes in the list, and you decide when something is finished.

What you need ​

  • An assistant that is connected to Kanbanto with Read and make changes.
  • An assistant that can keep going through several steps on its own. Claude can, in an ordinary chat: it works for as long as that conversation runs. Nothing works in the background while you are away unless you use a tool made for that.

Set up a queue ​

A queue is just a list. A common shape:

ListWho moves cards hereCounts as
For the agentYouNot started
DoingThe agent, when it starts oneIn progress
Your reviewThe agent, when it thinks it is finishedIn progress
DoneYou, after checkingDone

Add lists with Add another list, and set what each counts as from the list's ⋯ menu. "Your review" counting as in progress keeps a card from looking finished before you have looked.

Write cards an agent can act on ​

An agent only knows what the card says. A good card has:

  • a title that is one clear piece of work,
  • a description with what "finished" means, and anything it must not touch,
  • links to anything it needs. It can read text files and look at pictures attached to the card (a brief as Markdown, a screenshot). It cannot read a PDF or a spreadsheet: paste the text it needs into the description, or give it the file in the chat.

Put the cards in the order you want them done. The order of a list is your priority.

Tell the agent the rules, once ​

Paste a short instruction as your message in a chat with the assistant, for example:

Work through the cards in "For the agent" on my Website launch board, from the top. For each one: move it to Doing, do the work, comment on the card with what you did and how you checked it, then move it to Your review. If you are blocked or unsure, say so in a comment and leave the card where it is. Never move a card to Done.

If your assistant has a place to keep instructions between chats, put it there and you can then just say "work through my queue".

Where the work itself ends up. The assistant writes on cards: a comment, or the card's description. A draft, a summary or a list of findings goes there, or, when it is long, in a text file it attaches to the card (ask for that in your rules: "attach the report as report.md"). A document or a spreadsheet it makes in the chat stays in the chat with the assistant; it tells you so in its comment, and you download it from there.

To stop it, use the stop button in the chat. Cards it already moved stay where they are; the one it was working on is in Doing, with nothing lost.

What you do ​

  • Read the comment on each card in Your review.
  • If it is right, drag the card to Done.
  • If it is not, comment with what is missing and drag it back to For the agent.

Because everything is written on the cards, the next conversation with your agent starts where the last one ended. It does not need to remember: the board does.

Good to know ​

  • Break big cards down. Ask the agent to split a large card into subtasks first, and review that plan before it starts.
  • Use "Waiting on". If one card depends on another, open the card and pick the other one under Waiting on. Agents see that it is blocked.
  • Watch it live. Cards move on your screen as the agent works.
  • It cannot delete or share. A mistaken agent can move, edit or archive cards. Moved and archived cards are easy to put back. A description it rewrote can be put back too: open it full page and click Versions. Still, tell it to write its results as comments, as in the instruction above, and to leave descriptions alone unless you ask.
  • It waits for people. While someone is writing a card's description, an agent's change to that description is refused, and it is told who is writing.
  • Read-only is an option. If you only want answers, connect the assistant with Only read.

Next ​