The AI TODO list that does the items
Every other AI to-do app helps you organise, prioritise and schedule. Then you still do the work. Here the list is the interface to an agent: add "chase the three overdue invoices" and it logs into your accounting tool, sends the emails, and marks the item done with the receipts attached.
Organising a list is not the same as clearing it
Todoist, Motion, Reclaim, TickTick and the rest add AI to the list: natural-language entry, auto-scheduling, priority suggestions. Useful, and the work stays yours. TODO for AI starts from the other end. Each item is a task an agent can take, with a machine to run on, a browser that is logged in as you, and a model to think with.
That changes what belongs on the list. "Write the Q3 update", "fix the flaky test", "reply to the five support tickets", "post the launch thread" are all items here, because each one ends with a result and not a reminder.
What actually goes on the list
Real items from the six audiences we write guides for. One list, one agent, different jobs.
Chase overdue invoices. Draft and send the investor update. Fix the signup bug. Publish the comparison page.
Triage new issues. Bump the dependency and run CI. Review yesterday's PR with a second model. Write the changelog.
Pull each client's weekly ad numbers into one sheet. Update the CRM after the call. Ship the landing page copy.
Turn the blog post into five social posts and schedule them. Refresh the meta descriptions. Send the outreach batch.
Sort the inbox. Process the receipts folder. Answer the tier-one tickets. Prepare the monthly report.
Add, run, review, done
You add an item by typing or by voice. An agent picks it up with the permissions you set for it: allow, ask first, or block, per tool. It works on your PC, in your Chrome, or on a cloud VM, and posts progress into the item. When it is finished the item shows the result, the files, and the log. You approve, send it back with a note, or close it.
Items run in parallel, keep running when your laptop is closed, and can be scheduled.
Your colleagues and your agents, on one board
Invite people into a project and you get one board that your team and your agents both work on. Items carry the people who worked them and the agent that ran them, and the board filters by person, so what the company has in flight is one screen rather than a thread. An item can start as a note from your ops lead, get picked up by an agent overnight, and come back for a human to approve: one item the whole way through, not three tools and a handoff.
You are not running one assistant, either. Make as many agents as you have kinds of work, each with its own model, its own instructions, its own tools and its own permissions, each a different colour on the board. A careful one with approvals on everything for the billing project, a fast cheap one for triage, a coding one with a shell on your dev box. Items say which agent took them, and several run at once.
Companies do more than one kind of work, so they run more than one project: engineering, marketing, support, finance, one per client. Each project has its own members, its own agent, its own brand voice and its own cards. One team, one pooled budget, one place to look.
Items show their contributors and the agent that ran them, and the board filters down to one person, so you can see what is already in flight before starting it again.
Each agent has its own model, prompt, tools and permission set. Choose the one that takes an item, and run several in parallel.
Separate boards, members and agents for engineering, marketing, ops or per client, all under one team and one bill.
Read, write or owner per person, and allow, ask-first or block per tool, per agent. Set the tools that touch production to ask first and the approval is required, not remembered.
Members who can write to a shared board get a private card only they can see, so personal work lives on the same list without being published to the team.
After a while it adds items itself
The agent learns your business from what it has done: the site, the repos, the inbox, the accounts it has been inside. From that it proposes TODOs: the page that should exist, the invoice that is late, the dependency with a CVE, the review that has not been answered. You accept or dismiss. The list gets ahead of you instead of behind you.