Where Staleness Starts
Work goes stale in the gaps between touches. Every open item has a last-touched date, and the distance from that date to today grows each day. No single event marks the moment an item becomes stale. It happens because nobody looked.
Programs respond to this with review. Review comes in two forms, and each form fails in a different way.
Periodic Review
Periodic review is a scheduled pass over everything. A team sits down weekly, monthly, or quarterly and works through the open items. The pass is complete, since it covers every item. It forces prioritization, because time is limited. It also leaves a record of what was decided.
The weakness is the interval. An item that stalls the day after a review waits until the next one. Review sessions also drift into ritual: the same people read the same list and confirm the same statuses.
A periodic review works when it has a fixed agenda, a time limit, and a place to record decisions. Consider a hypothetical monthly review of open items. Without an agenda, the group discusses whatever is loudest. Without a time limit, the first few items consume the session. Without a record, the next review starts from memory.
Continuous Review
Continuous review puts something on watch all the time. That something can be an alert, a dashboard, or an agent. The blind spot is short, since a stalled item can surface within a day.
The weakness is volume. A constant stream of signals competes with other work for attention, and people mute streams they cannot act on. Ownership also blurs. When a system watches everything, each person can assume someone else acts on what it reports.
Using Both
The two forms answer different questions. Continuous review asks whether anything is stuck. Periodic review asks whether this is still the right work. The first question suits a daily report. The second needs people in a room with time to think.
My setup pairs a daily read-only report with a scheduled review that decides. The report shortens the blind spot between reviews, and the review uses the report as its starting material. Neither form replaces the other.
Where an Agent Fits
An agent can carry the continuous half. A report-only agent reads the tracker each morning and lists items that changed, items that went quiet, and items that look finished. It shortens the blind spot without asking for attention throughout the day, because the report arrives once.
The report ends at description. Each quiet item still needs a person to choose between asking the owner, changing the priority, and closing it. The agent supplies the list, and the periodic review supplies the judgment.
Cadence
A cadence sets how long an item can stall before someone sees it. I choose one by asking what a stall costs. Work with a deadline and downstream dependents gets a short interval. Work that changes slowly gets a long one.
Different programs carry different cadences, and each program's cadence is written next to its description. A cadence nobody wrote down defaults to whenever someone remembers.
Ownership
An item with a named owner and a next date can be checked. An item owned by "the team" cannot, since no individual answers for it. Reviews without owners produce lists. Reviews with owners produce decisions, because someone in the room is responsible for the next step.
When I audit a program for staleness, I count how many open items have both an owner and a next date. That count says more about the program's health than the number of open items does.
Reviewing the Review
A review that fails to happen leaves no trace, and a program can run for weeks on the belief that reviews are occurring. Someone needs to see that each scheduled review took place. A dated record works: the review produces a short note listing what it covered and what it decided. When the note is missing on the expected date, the missing note is the signal.
This applies to agents as well. A daily report that fails to arrive tells the reader something stopped, and the fixed schedule gives the reader a reason to notice.
Tooling
Tools store dates, send reminders, and display boards. Ownership and cadence come from people. Two teams that use the same tool with different cadences get different outcomes, and two teams with the same cadence and different tools get similar ones.
The order I follow is ownership first, cadence second, tooling third. The tool gets chosen to fit the first two. Starting with the tool reverses the order and leaves the ownership and cadence questions to whatever the tool suggests by default.
A Self-Check
Four questions apply to any program. Who owns each open item? When does each item next get looked at? What happens when the reviewer is away? What shows that the last review took place? A program that answers all four is unlikely to go stale unnoticed. A program that cannot answer one has found its first improvement.
The next post covers how Claude and I design agents that touch real work, and the limits we set on them.