Vol. 26 · Issue 11
Ben Zanghi — Notes & Work
Boston · Updated November 2026


A Backlog My Agents Can Write To

When I moved my agents' backlog to a tracker they could write to, an audit showed almost none of their planned work had ever been filed there. What I changed so agents and I share one list, and what a full disk showed me about running it myself.

A Backlog My Agents Can Write To

In October I moved the backlog for my personal AI agents to Plane, an open-source project tracker I run on my own server. I wanted one list that the agents and I both read and write, where anything waiting on me is obvious.

If your team is putting AI agents to work, the same questions come up quickly. Who did this? What's waiting on a person? And when something says "done," is it?

The backlog wasn't where the work was

Before switching, I had an agent audit how the work really happened. Over the previous three weeks, my agents had run 49 tracked tasks across about 30 projects. They asked me 22 questions and reported 53 blockers.

All of that lived on a live dashboard where agents post progress as they go. Only one of those 49 tasks had ever been filed as a tracker item.

Most of what was in the old tracker came from one assistant that files tasks on a schedule. It had written more than half the items. Every one of its writes, and every other agent's, went in under my name, so the history couldn't tell me who did what.

Status was split too. Items waiting on me carried ad-hoc tags like "needs-human" and "blocked." The tracker has built-in statuses for exactly those situations, and nothing was in them.

What I changed

Agents now write as a shared "agents" account, and the scheduled assistant has its own. Each machine has its own access key, and every key expires: 90 days for agent keys, one day for admin scripts. I can see who did what and revoke one machine without touching the others. The price is a rotation chore every few months.

Two statuses, "Needs Ben" and "Blocked," replaced the tags. "Needs Ben" is for a decision or step only I can take, and the item has to state the exact question and what the agent will do if I don't answer. "Blocked" means waiting on anything else. Each status now carries a short description, so an agent reads the rule where it works.

The plug-in that lets agents use the tracker came with 30 actions. Sixteen of them were for paid-edition features the self-hosted version doesn't include, and just describing all 30 used about 17,000 tokens of each agent's working memory every session. The trimmed version has the 14 that work, at about 8,000 tokens, with my conventions written into its instructions.

The move didn't replace the live dashboard. Agents still post progress there, and they can now link a task to its tracker item, so the plan and the progress point at each other.

Checking the move

The first migration check said 243 of 243 items matched. That was true for the export I'd checked against, but the old tracker kept receiving work for three more days, with 16 new issues and 60 comments.

Once the old tracker stopped receiving work, the final check counted 259 of 259 items and 1,091 of 1,091 comments, and I retired it on October 15.

The first day also showed the cost of the new status. My own tag-to-status conversion put at least one parked item into "Needs Ben" by mistake, and the scheduled assistant moved it back. By that evening 15 items sat in "Needs Ben," and nothing moved until I went through them. A waiting-on-a-person status is a queue, and someone has to work it.

A month in

Item numbers in my main project went from about 188 to 307 in the month after the move. Of the 148 items there that haven't been archived, 117 were filed by the agents' account, many of them at my request. The agents now log their own plans where I can see them, instead of only on the side dashboard.

The backup that said "ok"

In the evening of November 10 the tracker started returning errors. It shares a small server with another app I run. That app's database keeps a running log of changes and uploads it to cloud storage. When those uploads started failing, the log piled up on the local disk until the disk was full. The other app's database crashed, and the tracker stopped answering.

The cause was a storage limit on the backup account. Freeing some disk space brought the tracker back that night. To keep the other app running, I paused its uploads until I could raise the limit. They resumed about a day and a half later, after my second change to the limit actually took.

While fixing that, I found something worse. The tracker's own nightly backup used the same storage account. It had also been failing, while logging "backup ok" every night. There is roughly a week with no tracker backups at all.

The script is fixed, and old backups are now pruned so storage levels off instead of growing. I had tested a full restore twice during the move, and both passed. What I hadn't done was check that new backups kept arriving after that. There's still no alert for it, so for now I check by hand.

What I'd tell a team putting agents on a shared backlog

Give agents their own accounts. If every agent writes as a person, you lose the history and you can't revoke one without the others. Separate identities with expiring keys are worth the rotation chore.

Make "waiting on a person" a status with rules, and staff it. An exact question and a default turn "the agent is stuck" into something I can usually answer from my phone. The queue still needs someone to work through it.

Audit where the work happens before you migrate. My agents' real work was on a dashboard the tracker never saw. I kept both and linked them, which is less tidy and closer to how the work actually runs.

If you run it yourself, check that backups land. A log line that says "ok" is the job reporting on itself. Look for the new file, and alert when one is missing. That alert is the next thing I'm building.