Keeping an AI project backlog clean

Keeping an AI project backlog clean comes down to three rules: never delete a decision, merge duplicates as soon as you see them, and groom the backlog on a schedule rather than when it becomes unbearable. Agent projects collect clutter faster than ordinary ones. Agents suggest follow-ups at the end of every run, several sessions can file the same idea in different words, and tasks that were urgent last month sit untouched as stale TODOs. A cluttered backlog is not just untidy: agents and people start work from it, so a duplicate or an obsolete task can cause real wasted effort. This article covers how clutter forms, why declining and holding beat deleting, how to handle duplicates and ID collisions, and a monthly grooming routine. If you are seeing agents redo finished work, the troubleshooting page on repeated work covers that symptom directly.

How agent projects accumulate clutter

Four sources account for most of it. Follow-up suggestions: at the end of a task, agents often list improvements they noticed. Captured uncritically, each becomes a task, and few are ever ranked. Parallel sessions: two agents, or the same agent on different days, file the same issue as fix rounding in export and CSV totals off by a cent. Abandoned ideas: tasks created during an exciting planning session that nobody returns to. Superseded work: tasks that made sense before a refactor and now describe code that no longer exists.

The cost is not only visual. A backlog that agents or scripts read from can feed them an obsolete task with full confidence. A duplicate can be worked twice. A stale TODO can be picked up and implemented against assumptions that stopped being true weeks ago. And people stop trusting a long backlog, so they keep private lists, which defeats the purpose of having one.

The goal of hygiene is a backlog where every open item is something you would still agree to do if it were proposed today.

Decline or hold instead of deleting

Deleting a task removes the record that someone once thought it mattered and that you decided otherwise. Without that record, the same idea comes back, from you in three months or from an agent tomorrow, and you spend time deciding again.

Use two explicit states instead. Declined, often called WONT-DO, means you considered the work and chose not to do it. Add one line saying why: superseded by the new export module, or not worth the review effort for one user. Held, often called ON-HOLD, means the work is valid but paused on purpose, for example until a dependency ships or a client confirms. Add the condition that would bring it back.

Both states keep the task searchable and out of the active backlog. When the idea resurfaces, you find the earlier decision and its reason in seconds. Holds also deserve a review date, because a hold with no condition and no date is just a stale TODO with a better name.

Delete only true junk: test tasks, accidental empty cards, or exact copies created by a glitch.

Duplicates and ID collisions

Duplicates are rarely word-for-word copies. Agents describe the same problem in different language, so matching titles is not enough. When grooming, sort or search by the area of code a task touches, such as a file, module or feature, and read neighbours together. When you find a pair, keep the one with the better description and history, merge any useful detail into it, and close the other as a duplicate with a link to the survivor.

Prevention helps more than cleanup. Before creating a task, search for the file or feature it concerns. Ask agents to list candidate follow-ups in their report rather than creating tasks directly, so a person can check for existing ones first.

ID collisions are a separate problem. They happen when two sessions create tasks at the same time and both take the next number, or when you import tasks from another project that used the same prefix. Use one place that issues IDs, use distinct prefixes per project, and preview any import before applying it so conflicts are visible rather than silently merged or overwritten.

A monthly grooming pass

Step 1: Sort open TODO and BACKLOG items by last update, oldest first.

Step 2: For each item untouched for a month, ask whether you would accept it if proposed today. If yes, confirm its priority and description. If not, decline it with a reason, or hold it with a condition and a date.

Step 3: Review every hold. Has its condition been met? If so, return it to the backlog. If the condition is no longer relevant, decline it.

Step 4: Search by area for duplicates, focusing on modules with heavy recent agent activity.

Step 5: Check that tasks still match the code. Anything that names a removed file or function needs rewriting or declining.

Step 6: Re-rank what remains. See prioritizing when agents work fast.

The pass should take under an hour for a typical small project. If it takes much longer, groom more often or capture fewer follow-ups to begin with. Stale tasks are a form of stale project memory, and they mislead agents in the same way.

QUESTIONS

Should I delete old tasks from my backlog or close them?

Close them with a reason rather than deleting them. Mark work you decided against as declined and paused work as on hold, each with one line explaining why. The record stops the same idea returning later as if it were new. Delete only test entries and accidental copies.

How do I find duplicate tasks created by different AI agents?

Search by the file, module or feature each task touches rather than by title, because agents describe the same problem in different words. Read neighbouring tasks together. Merge useful detail into the better task and close the other as a duplicate with a reference to the one you kept.

How often should I groom a backlog on an agent-heavy project?

Monthly is a good default, with a quick prune of new low-priority items in your weekly planning. Projects where agents capture many follow-ups may need a lighter pass every two weeks. If grooming regularly takes more than an hour, capture fewer follow-ups at the source.

RELATED GUIDES

• How to find and organize work across projects
• How to create a project card in BRAIN FLUID Desk

RELATED HARNESS PAGES

• Why does my AI coding agent keep repeating work? Agents repeat work when nothing records what was done or declined, tasks are duplicated, or context is lost. How to tell which, and fix it.
• Prioritizing a backlog when agents work fast Prioritizing a backlog when agents work fast: rank by value, review effort, risk and dependency order, and use P0 to P3 as real commitments.
• Managing tasks an AI agent discovers mid-work Managing tasks an AI agent discovers mid-work: record each as a new linked task, mark it required or optional and never widen the current task silently.
• Stale AI project memory Stale AI project memory: how notes and evidence an agent reads fall out of date, how to detect it, and what a freshness check cannot tell you.
• kanban for AI agents Kanban for AI agents: the board states that matter when software does the work, including review, blocked and decision owed.