Delete Your Backlog? Try an Agent Burndown First

I used to believe in deleting your product backlog every once in a while. But now, agents can dig through it and not only clear the deck but also find and advance the hidden gems for you.

It’s still true that if it's important, it'll come up again. Lots of codebases are moving faster and rendering old things irrelevant.

But, there might also be a key insight buried in there that wasn't the most urgent thing at the time but that is still a big win. Or maybe fixes the thing that has been annoying a good customer just enough to look for another option.

The Experiment

In mid-June, my client RootNote had 588 issues in the backlog that had built up over the course of a year or so. Around that time I started thinking about a fresh start.

At the time, I had really shifted to thinking about all the things I can pass off to agents. I thought “instead of nuking it from orbit, what about a controlled burn?"

Every two hours on weekdays, the automation:

• Fetches Linear issues with a "Todo" or "Backlog” status.
• Excludes anything it has already reviewed - each one is tagged.
• Selects the oldest issue within the highest available priority.
• Checks it for relevance against the current codebase.
• If it's still relevant, and has enough info to implement with 90% confidence, it automatically triggers my chain of skills that carry an item through implementation, iterative review, and direct testing before a final human review and test pass.
• If there's not enough information, it asks clarifying questions on the linear issue or in the running agent thread that can be answered in a batch along with its recommendations and trade-offs to consider.
• If it's not still relevant - it recommends closure or finds the duplicate issue.

I cap the maximum number of automated cleanup PRs at 10 - I don't want to overwhelm the system with changes that aren't making it all the way through the process. It's also why it doesn't run over the weekend. That and I don’t want the temptation to look at it.

The Numbers

From June 18th to August 12th, here are the stats:

• 407 issues reviewed
• 239 closed out:
  • 114 Moved through the process to done + deployed
  • 60 canceled as irrelevant
  • 65 linked as duplicate to what was already solved
• 45 in delivery or testing
• 123 refined and still queued
• Todo + Backlog: 477 → 344 since July 13
• Net reduction: 133 issues, or 28%

The net reduction includes all work entering and leaving the backlog via the team, the robots, and me.

Still a big improvement!

Some Things the Robots Did

Beyond the stats, here are a few examples of meaningful improvements that got shipped because this process existed that I wouldn’t have remembered or spent the time to dig up myself:

Hide accidentally synced content

A RootNote user is typically a content creator or an agency working with one. Odds are, they have multiple accounts for Instagram or TikTok, etc. to choose between when connecting.

We don’t control the whole connection process, the platforms do. Given that, it’s hard to make the process completely seamless, and people inevitably connect the wrong account to their creator profile.

When this happened, we were manually cleaning it up via support requests - but it didn’t happen that often.

The cleanup agent found this backlog item that was 11 months old, ran it through our feature pipeline, and made some pretty good design decisions after it asked me a few questions.

Now, users can bulk “archive” content from the wrong account, and restore it if needed from an easily accessible screen.

And we never have to clean it up manually again.

RootNote’s hidden content screen for restoring content archived from the wrong connected account.

Automatically focus the search box

Tagging synced content is a key workflow of RootNote, and there’s an easy way to select one or more pieces of content and add or create a new tag. One big annoyance - the tag search field wasn’t automatically being focused when it opened.

This four-month-old issue actually had been in the codebase since the tag search was implemented. I had tried a year ago to fix it but I couldn’t untangle the React maze that was causing it to happen. But the agent ripped through it, verified in a browser it was working all on its own, and served me up a heavily reviewed and hardened PR before I did my own review and test.

One less paper cut for a key workflow.

RootNote’s tag search workflow after the search field is automatically focused.

Connect Data to Copied Dashboards

A core feature of RootNote is “Dataspaces” - easy to create dashboards that show off your data for prospective brand partners and collaborators that always stay up to date.

We had introduced the ability to copy them, as well as copy a template created by the RootNote team, but for some technical reasons at the time automatically connecting the right data didn’t make the release. People copying a dataspace had to manually connect each data source, making it save a lot less time than they expected it to.

6 months later, it was no problem for GPT 5.6 running inside my delivery pipeline setup. Now it works as it was always supposed to.

RootNote’s Dataspace dashboard showing the data connections available from the overflow menu.

Closed Out Things that Weren’t Relevant Anymore

When a need pops up, we capture it. Whether that’s coming from users, internally, or is thrown off as a follow up issue as part of my delivery pipeline. That’s something that happens often.

Especially before this automation existed, it was easy for things that weren’t urgent to fall through the cracks, and I’d continue on my merry way improving the architecture and the product.

As a result, there were a lot of captured items - especially follow ups capturing tech debt - that I naturally either solved or made irrelevant while working on a bigger picture effort like migrating how we do syncs to something simpler and easier to manage.

The backlog cleanup agent took a look at what was scoped - reconciled it against the code, noticed it was an irrelevant change and recommended it be closed.

Found Duplicate items

Just like making sure it’s still relevant, making sure it’s not already done is a key job when both humans and agents are capturing every todo item in your product.

If it finds the specified change implemented already, it recommends closing the issue as a duplicate of the one that already completed the work.

The Takeaway

It may not make sense for you to build this exact workflow. But, you might have a big pile of stuff that has some valuable things in it.

Agents directed well can help you find the good stuff and clear out the rest without worrying if you're letting something slip that you shouldn't.

When you have work that’s important to get done but not urgent so it just keeps slipping, hiring an external partner can be a good solution. But most of the time it just sits undone or gets written off.

Now, you can consider a more accessible, cheaper option - hire an agent to do it, at a pace you can absorb without having to think too hard about it.