A decision can be made in one place while an older version remains somewhere else. A document still describes a proposal. A task list still shows work that has finished. An implementation has changed, but the explanation of it hasn't caught up.
I've been experimenting with using an agent to compare those records and bring the differences back for review. I think of it as a drift watch: a way to notice when the account of the work and the work itself start moving apart.
What I've been trying
The approach overlaps with my weekly priorities review. Instead of carrying the last list forward unchanged, look for evidence of what has happened since: a reply, an accepted decision, a merged change or a result that can actually be inspected.
This matters across tools too. Work may happen in Claude Code while the brief or priorities review lives elsewhere. A session summary helps with the handoff, but I still want the underlying change and checks. Access to one assistant's conversation doesn't imply access to another's.
Then separate the states. Prepared isn't sent. Agreed isn't implemented. A merged change isn't necessarily live. Keeping those distinctions visible makes the next action much clearer.
I've been applying the same question across documents, definitions and implementation records: do these still describe the same thing? Where they don't, the useful output is a specific discrepancy with the sources beside it, ready for someone to assess.
What I'm learning
The difficult part is interpreting the difference. An old note may be deliberately historical. A newer message may be a suggestion rather than a decision. An empty result may mean the check couldn't see the source, rather than that something has disappeared.
For example, if a check returns no records, I want the next question to be whether the check worked. It shouldn't jump straight to diagnosing the system it was supposed to inspect.
That changes what I want from the agent. Show what it read, what it couldn't read and why it thinks the difference matters. Keep the proposed correction separate from the evidence supporting it.
Where this stands
This is still a working approach, including manual catch-up and review. It isn't a claim that a background service reliably catches every change, or that all the connected systems have been audited.
The next test is whether the comparisons consistently surface useful corrections without repeatedly reopening completed work or creating noise from missing information. Some differences should prompt a change; others should remain an explicit uncertainty.
If this works well, it should make the existing records more trustworthy and reduce the effort of remembering how they fit together. I don't particularly need another dashboard to check.