Forty Trello boards across a team is the number where things break. Half are alive but slow. A quarter are dead. Another quarter are alive but unowned, with cards that have not been touched since someone's last role change. None of them are obviously bad on the day they were created. They drift there, one weekly standup at a time, and after eighteen months the team is paying a tax that no single person can see.
The cost of dead-but-visible boards is not the storage. Trello does not bill you for cold cards. The cost is the cognitive overhead of every Friday review, every "where did we land on X" search, every onboarding that takes a new hire through six boards before they find the one that is actually used. Visual noise on a kanban surface is a form of debt. An AI agent that runs once a week, archives the corpses, and tells you which living boards are sick is one of the cheapest ways to pay that debt down.
What this agent does
This is a hygiene agent, not a card-creation agent. It does not write new cards. It does not assign owners. It does not move cards between lists. It does one thing: it reads every board you give it access to, finds the defects, and proposes a clean-up.
Once a week, on the schedule you set (Friday afternoon is the default), the agent walks every board, scores it on four hygiene defects, generates an archive proposal, and pushes a single digest to the channel you pick. You open the digest, glance at the board-health scores, and approve the archive groups you trust. Nothing happens without that approval.
For the broader read-then-write pattern that this agent sits inside, see what an AI agent can actually do.
Sources of truth
The Trello REST API is the one source. The agent reads three things and writes one.
- Boards and lists.
GET /1/members/me/boardsandGET /1/boards/{id}/lists. The shape of the workspace. - Cards and their action log.
GET /1/boards/{id}/cardsandGET /1/cards/{id}/actions. Activity history is the staleness signal. - Members and labels.
GET /1/boards/{id}/members. Used for the ownership check. - Write path: archive only.
PUT /1/cards/{id}/closedwithvalue=true. That is the only write the agent ever issues.
The agent does not read your email, your calendar, or any system outside Trello. The boundary is the workspace. This is on purpose. The fewer integrations the agent has, the smaller the blast radius when something goes wrong. For why that matters, see how to limit agent actions.
The four hygiene defects
The agent looks for exactly four things. Anything else is out of scope and pushed to a future version. Resist the urge to make this a do-everything board manager. Hygiene agents earn trust by being narrow.
Stale cards. Any card with no activity for 60+ days. Activity is anything in the card action log: comments, member changes, due-date edits, list moves, checklist updates, attachments. A card that is two years old but had a comment last week is not stale. A card created last month with no activity since day one is. The 60-day default is a starting point; some teams need 30, some 90.
Stuck WIP. A card sitting in a list labelled In Progress, Doing, Working, or any custom WIP list you flag, for more than 30 days. The classic kanban anti-pattern. The Atlassian Agile Coach guidance on WIP limits is the source for this: work that stays in progress past a reasonable cycle is signalling a blocker that nobody is naming.
Unowned high-priority cards. A card with a high-priority label (red, the High Priority label, or any colour or label you flag in config) and no member assigned. Or a card with a member assigned but the member has not been active on that board for 30 days. These are the cards that quietly become someone else's problem.
Redundant boards. Boards whose names and card titles overlap heavily with another board. The agent uses a simple lexical similarity score (Jaccard on tokens of board title + top card titles) and flags pairs above a threshold. It does not propose merges; it asks you whether the pair should be merged. This is the lowest-confidence defect of the four, so the agent only ever reports it.
Output: weekly digest and archive proposal
One message per week to a Slack channel or an email inbox. The structure:
- Board-health scoreboard. One row per board: active cards, stale cards, stuck WIP, unowned high-priority. The numbers, not adjectives.
- Archive proposal, grouped by board. Each group has a single Approve archive button. You see the count and a few sample card titles, not all 47 cards.
- Stuck-WIP cards. Card title, owner, days in WIP. No action button. The point is to surface the blocker, not to move the card.
- Unowned high-priority cards. Card title, board, label. Action button to ping the board admin.
- Redundant-board suggestions. Pairs of boards above the similarity threshold. Read-only. The human decides.
The whole digest should be skim-readable in under five minutes for a 40-board team. If the digest takes longer than that to read, you have too many defects per board, which is the actual problem the agent is supposed to make visible.
Guardrails
- Archive, never delete. The agent does not have access to the delete endpoint. By design.
- Protected boards are read-only. The main release-track board and any board you flag are scanned and reported, but no archive is ever proposed for cards on them.
- No auto-archive of cards with recent comments. Even if a card is past the staleness window, if a comment landed in the last 7 days the card is excluded. Comments are the strongest signal of latent activity.
- Humans approve archives in bulk. No silent archives, ever. The approve action is per-board, not per-card, to keep the digest fast to action.
- Rate-limit the writes. Trello caps API calls at 300 requests every 10 seconds per API key, per token. The agent's archive batch must respect this; on a large workspace, archiving 500 cards is a multi-minute job.
- Rollback is one click. Every approved archive batch is logged with the card IDs, so the human can issue a single restore command if the wrong list was approved.
The two posts on the patterns behind these rules: AI agent safety and guardrails and how to add a human approval step to an agent. For the rollback piece, how to roll back an agent action.
Common mistakes
- Archiving noisy-but-loved cards. Some cards are reference material that nobody comments on but everybody opens. The agent does not see opens via the API, so the recent-comment exclusion is the closest proxy. Tune the staleness window up if your team uses Trello as a knowledge base.
- Ignoring labels. A team that has spent a year setting up label conventions does not want the agent treating every card the same. Labels feed the priority defect and should also be readable in the digest so a human can spot a mis-categorised card before approving the archive.
- Treating colour as gospel for priority. Red does not mean High everywhere. Some teams use red for blocked. Some use red for external. Read the label text, not just the colour, and let the config override the default.
- Running the agent on day one with full write access. The first month, the agent should run in dry-run mode: it generates the digest but the Approve archive button is greyed out. You read the proposals, eyeball whether they are sensible, and only after a few cycles do you let it write. The same graduation pattern as every other agent. See how to test an agent before deploy.
- Forgetting to surface stuck WIP as a leading indicator. The stale-card count is lagging. Stuck WIP is leading. A team with a rising stuck-WIP count this week will have a rising stale count next month. The digest should treat stuck WIP as the headline, not the footer.
- No monitoring on the agent itself. A hygiene agent that silently breaks for three weeks is worse than no agent. Log every run, fail loudly. See how to monitor agent activity.
Frequently asked questions
Can an AI agent archive Trello cards for me?
Yes, but the right framing is hygiene, not deletion. The agent runs once a week, scans every board you give it access to, and produces a list of candidates: stale cards with no activity for 60 days, cards stuck in a WIP list for 30 days, and high-priority cards without an owner. You approve the archive list in bulk. Trello's archive is reversible, so a wrong call is recoverable, but the human is still in the loop for the actual archive action.
Why archive instead of delete?
Trello archive is a soft action. The card is removed from the board view but still searchable and restorable from the board menu. Delete is permanent and irreversible. An AI agent should default to archive every time. If a card needs to be deleted, that is a human decision made later from the archived list, not an autonomous one. This single design choice is what makes a Trello hygiene agent safe to run weekly.
What counts as a stale card?
The defensible default: no activity on the card for 60 days. Activity means any change in the card's action log via the Trello API, including comments, member changes, due-date edits, list moves, checklist updates, or attachment additions. The agent should not infer staleness from card age alone. A card that has sat on a Someday list for two years but had a comment last week is not stale; it is parked.
How does the agent know which board is the main release-track board?
You tell it. The agent reads a config that flags one or more boards as protected, and on protected boards the agent only flags and reports; it never proposes archives. This is the same pattern as protected branches in source control. Without this rule, a Trello hygiene agent will eventually try to archive a card that is mid-release and create a real incident. Protect first, scan second.
What does the weekly board-health report look like?
One email or Slack message every Friday. Top of the message: a board-health score per board (active cards, stale cards, stuck WIP, unowned priorities). Below: the proposed archive list, grouped by board, with a single approve button per group. Below that: stuck-WIP cards with their owners pinged. Below that: unowned high-priority cards. The whole thing should be readable in under five minutes, even for someone with 40 boards.
Three takeaways before you close this tab
- Hygiene, not deletion. Archive is reversible. The agent never deletes.
- Four defects, weekly cadence. Stale, stuck, unowned, redundant. Every Friday.
- Protect the release board. One config flag prevents the one bad day.
Sources
- Atlassian Developer, "Trello REST API: Cards and Actions", retrieved 2026-05-12, developer.atlassian.com/cloud/trello/rest/api-group-cards
- Atlassian Developer, "Trello REST API: Boards", retrieved 2026-05-12, developer.atlassian.com/cloud/trello/rest/api-group-boards
- Atlassian Agile Coach, "Kanban work in progress (WIP) limits", retrieved 2026-05-12, atlassian.com/agile/kanban/wip-limits
- Nielsen Norman Group, "Visual Hierarchy and Visual Noise on Dashboards", retrieved 2026-05-12, nngroup.com/articles/visual-hierarchy
- Aryan Agarwal, "Gravity Trello hygiene agent spec", internal v1, May 2026, About