Stop updating your task board by hand: let the agent move the card

You stop updating a task board by hand when the agent doing the work moves the card as part of doing it. In Weft, start_task puts the task in Doing, report_progress keeps the claim alive while the agent works, and complete_task moves it to Done, or to your Inbox when the task needs your review. You stop being the person who keeps status true and become the person who decides.

We make Weft. Facts about other products come from their official pages, checked on 27 October 2026.

Why does every task board go stale?

Status is a second job nobody is paid for. The work is writing the code, sending the invoice, fixing the page. Moving the card is overhead, and it comes right when you want to move on.

So the card gets moved later, in batches, from memory. By Thursday, Doing holds three things nobody is doing and Todo holds one thing that shipped on Monday. People stop trusting the board, check chat instead, and the board turns into a picture of last week.

AI agents make this worse. An agent can finish five tasks in an afternoon, and if you still drag each card yourself, you do clerical work for a faster worker.

Is the fix cleanup or prevention?

There are two ways to automate a board. Cleanup lets the drift happen and then uses AI to sort it out. Hypertask's article Stop Manually Updating Your Task Board takes this route: it describes an AI Triage button that reads your inbox, labels each item safe to archive or needs your input, and archives the first group once you approve. If your pain is a flood of comments and pings, that is a reasonable tool, and Hypertask has built-in AI for it. Weft has no built-in AI and no notification triage.

Prevention is the other route: the status never goes wrong in the first place, because changing it is not a separate step. The agent cannot start the work without claiming the card, and it cannot finish without closing it. There is nothing to triage later, because the card already says what happened.

We chose prevention. A board that needs a cleanup pass was wrong between the passes.

How does the agent move the card while it works?

Weft connects to your AI client over the Model Context Protocol at https://letsweft.com/api/mcp. The client signs in with OAuth in the browser, so there is no key to copy. Four tool calls carry the status of a task from start to finish:

1. start_task claims the task and moves it to Doing. The claim is a 20-minute lease. If a second agent tries the same task, it gets already_claimed and is told not to duplicate the work.
2. report_progress is a heartbeat. Each call renews the lease, and the response tells the agent to renew every 5 minutes. If the agent crashes or the chat is closed, the heartbeats stop, the lease lapses after 20 minutes, and the task is freed for anyone to pick up. Silence is a status too.
3. request_input asks you a question, with options, while the agent keeps its claim. The question lands in your Inbox and can reach your phone as a push notification.
4. complete_task records the outcome, the artifacts that prove it (a URL, a commit, a file hash), what was not done, what is still unknown, and the next step. By default the task moves straight to Done. If the task is marked as needing review, or a verification check fails, it stops in your Inbox instead.

The agent does not need to remember to update anything, because the update is the call it makes to do the work. For the full mechanics of leases and receipts, see the claim loop guide.

One honest limit: Weft never calls an agent. The agent pulls work when you are in a session with it, through get_my_work or start_task. Cards move because an agent you are running moves them, not because Weft watches your repository. Weft has no GitHub or Slack sync and no webhooks, so a pull request merged outside an agent session will not move a card by itself.

What does a manual board look like next to an agent-moved one?

What should I automate, and what must stay my decision?

Automate anything that is a fact about the work: started, still going, finished, stuck. An agent knows these better than you do, because it is the one doing the work. Keep anything that is a judgement about the work: is this good enough, should it ship, is this still worth doing.

In Weft the right-hand column has two doors. When the agent hits a decision mid-task, it calls request_input and you answer one question, ideally by picking an option. When the agent finishes work that needs your review, you see four buttons: Accept, Rework, Later, Drop. Rework asks you for a note. If the agent passed a next step, you see it above the Accept button, so accepting approves exactly that step and nothing more.

Can the AI set up the board from a conversation?

Yes, and it is the easy part. Describe the project in your chat, and the agent can call initialize_board to create the four default columns (Backlog, Todo, Doing, Done) and then bulk_create_tasks to add the whole plan in one transaction. The batch either lands in full or not at all, and it counts as a single call against your monthly quota. A task created without a column lands in Backlog.

Hypertask's post Why We Let AI Agents Build Entire Boards argues that the project, not the single task, is the right unit to hand to an agent. We agree that agents should build boards. We disagree that building is the hard part.

Building a board is cheap: an agent can write forty tasks in seconds. Reviewing is the real limit. Each of those tasks will one day produce finished work you have to look at, and you can look at only so much in a day.

That is why Weft has an attention limit instead of a creation limit. When too much is waiting on you, start_task refuses to claim new work and answers attention_queue_full, along with what the agent may still do (report progress, check its work, read the board). The limit is 5 items for a one-person workspace and 10 for a team, or 45 minutes of estimated review time, whichever comes first. Questions from request_input count toward it too, because answering them costs the same attention.

An agent can plan as much as it likes, but it cannot start the sixth thing while five finished things wait for you.

What should my first week look like?

Connect one client first. In ChatGPT, Weft is a published app on every plan, including Free. In Claude.ai or Claude Desktop you add it as a custom connector. In Claude Code it is one command: claude mcp add --transport http weft https://letsweft.com/api/mcp (see the Claude Code setup page). Then paste this into the chat and edit the project line:

You are connected to my Weft board. For this week, follow these rules.

Setup (today):
1. Call get_board_state. If there are no columns, call initialize_board.
2. My project: [describe it in 3-5 sentences].
   Propose at most 12 tasks as a list. Show me the list first.
   After I say yes, add them with one bulk_create_tasks call.

Every time you work on a task:
3. Call get_my_work to see what is yours and whether my queue has room.
4. Call start_task the moment you begin real work, not while we are
   only discussing it. Keep the runId and leaseToken.
5. Call report_progress every few minutes while you work.
6. If you reach a decision that is mine (money, publishing, anything
   irreversible, or a real fork), call request_input with 2-4 options.
   Do not guess and do not ask permission to continue.
7. Finish with complete_task: a one-line outcome, artifacts I can
   check (links, commits, files), notDone, unknowns, and a nextStep.
8. Never move a card to Done by hand. complete_task does that.
9. If start_task says attention_queue_full, stop starting new work.
   Tell me what is waiting for me and help me clear it.

At the end of each session, tell me which cards you moved and why.

Twelve tasks is deliberate. On the Free plan you have 50 active tasks and 200 AI calls a month, and a first week is about learning the loop, not filling the board. If you are moving an existing board, Weft imports CSV from Linear, Jira or Notion and JSON from Trello on every plan.

During the week, do only two things: answer questions when they arrive and make a call on what reaches your Inbox. Do not drag cards. If a card ends up in the wrong place, that is worth knowing: tell the agent which rule it skipped instead of fixing the card quietly.

By Friday, check two things: does every Done card have something you could open to check it, and did anything reach you that you could have decided in advance? If so, write that default into next week's prompt.

What does Weft not do here?

• It has no built-in AI. The thinking happens in your own AI client, so Weft has no AI credits to buy and no triage button.
• It does not move cards from outside events. No GitHub, GitLab or Slack sync, no webhooks, no public REST API. If your status should follow commits and merges with no agent involved, a tool with those integrations fits better.
• It does not make every task wait for you. Most finished work goes straight to Done with its receipt. Only tasks that need review, or fail a check, stop in your Inbox.

If you run several agents at once, the claim is what keeps them apart; one board for multiple agents covers that case, and what did my AI agent do covers reading the receipts. For a wider comparison of task managers with MCP, see the MCP task manager table.

If the agent moves cards, how do I know it did not just mark everything done?

Done in Weft comes from complete_task, which records the outcome, artifacts, what was not done and what is unknown. Read the artifacts, not the summary. For work you want to check before it counts, mark the task as needing review and it will stop in your Inbox for Accept, Rework, Later or Drop.

What happens when the agent crashes halfway through?

It stops sending report_progress. After 20 minutes without a heartbeat the lease lapses and the task is freed for anyone, including another agent, to pick up. You do not have to notice the crash for the board to stop claiming someone is working on it.

Will the attention limit slow my agents down?

It slows down starting, not finishing. When 5 items (10 for a team) or 45 minutes of review are waiting on you, start_task refuses new work, but the agent can still report progress and read the board. The fastest way to unblock it is to make a call on what is already in your Inbox.

Can I keep some agents read-only?

Yes. Each OAuth connection gets full, write or read rights, chosen on the consent screen or later in Settings. A read connection can look at the board but cannot move cards. You can also disconnect a client at any time.

Do I need a paid plan for this?

No. The whole loop works on Free, with 50 active tasks and 200 AI calls a month; bulk_create_tasks counts as one call. Pro is $7 per seat a month or $60 per seat a year and removes those caps.

Is this better than AI triage?

It solves a different problem. Triage cleans up a feed of comments and notifications after the fact; if that feed is your problem, a tool with built-in triage, such as Hypertask, fits better. An agent-moved board keeps the status from drifting in the first place, for work your agents do.