If you’ve started using AI agents for real work, you’ve probably noticed they’re good at doing things and terrible at telling you what they need from you. This piece is for anyone who has hit that wall. It’s how I turned my own knowledge base into an AI task manager that my agents and I share, and what I learned in the first week.
The short answer to the question I kept asking myself: yes, AI agents and a person can share one task list. You need a task manager with an API or an MCP server, a tag for each agent, and one tag that means “a person has to answer this”.
Too many agents, too many files
I run a handful of Claude Code sessions for my business. One handles outreach, one content, one PR, one a build we’re doing with a colleague, and since last Thursday one markets Ultrathink, the knowledge base I built for my own scattered brain. Each has helpers it hands work to.
Until last Tuesday, each session kept its tasks in its own markdown file. Long, tidy, carefully updated files. I hardly ever opened them.
That was the problem. The work was getting done, but the things that needed me (a decision, an approval, a “who should we email about this?”) were buried in five files I didn’t read, and I only saw them when I went looking.
Can AI agents and a person share one task list?
They can, and it’s simpler than it sounds.
The tools have caught up. Todoist, Asana, Linear, Notion, ClickUp and Taskade all now have an MCP server or an API that an outside AI agent can use to create and update tasks. MCP, the Model Context Protocol, is the open standard Claude, ChatGPT, Cursor and others use to connect to other software. Anthropic’s own docs put it well: connect a server “when you find yourself copying data into chat from another tool”, and Claude can then “read and act on that system directly”.
Claude Code does have a task list of its own, but Anthropic describes it as “Claude’s to-do checklist”: the steps it plans for its own work. It’s useful, but it’s Claude’s list, on your machine. It isn’t where you would keep your own tasks, decisions and questions.
So the real question isn’t whether agents can write to a task manager. It’s how to set it up so you only see what matters to you.
How I set it up
On 6 October I moved every session’s task list into Ultrathink. This is how it works.
One list, tagged by stream. Every task, decision, idea and question an agent creates goes into Ultrathink through its REST API, with a tag for the stream it belongs to: stream-outreach, stream-stories and so on. I can look at one stream, or all of them.
One tag that means “Chris, you’re needed”. Anything that needs me gets the tag needs-chris, written as a question I can answer. That one filter is pretty much my inbox now. When I answer, the agent removes the tag and carries on.
Tasks and knowledge kept apart. A task has a status (not started, in progress or done) and a due date only when there’s a real deadline. A decision or an idea is stored as knowledge instead, so it doesn’t clutter the list but can still be found later.
A name on everything. There’s no owner field, so the last line of every entry says who owns it: me, a colleague or a named agent. Small thing, but it stops tasks falling between two agents.
Waiting has its own tag. When an agent is waiting on a colleague, it tags the task “waiting” and names the person. So I can see what’s stuck and who it’s stuck on.
It survives the internet dropping. If Ultrathink can’t be reached, the little script the agents use queues the write and sends it next time. Nothing gets lost when my laptop moves between office and home.
What happened in the first week
About 540 entries went in during the first few days, most of them the old markdown notes moving across. Then it settled into a steady stream of new tasks and questions.
Three things surprised me.
The AI pass helps and occasionally embarrasses itself. Ultrathink runs an AI pass on every entry: it fixes spelling, writes a summary and adds its own tags. On agent entries that’s mostly useful. It also tagged a task about reading our website analytics as “web server”.
Agents over-explain. Their notes ran long. The fix was a rule: the title is the thing itself, in one line, and the question for me goes in the first line of the notes.
One place to look changes the habit. The needs-chris filter is the first thing I check now, and the calendar view shows everything with a date on it, from blog posts to approvals.
What to look for in an AI task manager your agents can use
If you want to try this, these are the things that mattered more than any feature list:
- An API or MCP server with write access, and keys you can scope per agent. Ultrathink’s keys can be read only, write only or both, and the API needs a Pro plan.
- Tags or labels you control, so you can give each agent a stream and agree your own “needs me” tag.
- Filters you can save and open on your phone. If the inbox isn’t quick to check, you won’t check it.
- Somewhere for decisions, not just tasks. Half of what agents produce is “we decided X”, and it needs a home that isn’t your to-do list.
- Due dates and a calendar view, so dated work shows up without you hunting for it.
Any of the tools above can do most of this. I use Ultrathink because the tasks sit next to everything else I capture: the links, notes and AI chats the work came from. When an agent asks me a question, the context is usually one search away.
What I’d do differently
I’d write the rules for the agents first. Which tags, what a good title looks like, when something needs a person and when it doesn’t. I made most of those up as I went, and the first couple of days were noisy because of it.
I’d also keep a person in charge of anything that leaves the building. My agents can draft an email or a post, and they can track it, but nothing goes out until I’ve approved the exact text. The task list makes that easier, because the approval is just another needs-chris question.
If you’re running more than one AI agent and finding out about their questions too late, try giving them one list and one tag that means “a person is needed”. It’s made a real difference to how I work with them.
