What I found was a second governance system. The workspace had grown its own, built independently of the vault documents I had already declared authoritative. Mission Control, an API on the agent host, was its declared system of record for task status, activity, and deliverable evidence. A "no invisible work" rule put task IDs and cost tracking on all agent work. There was a written confidential-data policy defining storage boundaries, a repository map, and a worktree-and-PR development procedure. All of it operational. None of it aware of the vault.
Meanwhile the vault, my notes system, versioned in a private Git repo, held the decision log, the source-of-truth matrix, and the governance rules I had written as the official layer. So I had two authoritative task systems and two source-of-truth maps that did not know about each other. Each internally consistent. Each blind to the other.
This was not a visible production failure. It was an inspection finding during a governance review: the decision log says, "This created two authoritative task systems and two source-of-truth maps that did not know about each other." The answer was the sentence that kept both honest: "No migration of a working operational system; the two maps are reconciled by reference, not merger."
The obvious fix was the wrong fix
The instinct in that moment is to consolidate: pick a winner, migrate everything, declare one true system. I decided against it, and the reasoning is worth recording.
The workspace governance was not a mistake to correct. It was a working operational system that had earned its rules through daily use. Migrating a functioning system so a document can claim completeness is backwards; the document should describe reality, not the other way around.
The actual problem was narrower: the two systems answered overlapping questions with no rule for which one was right. That is not a migration problem. That is a layering problem.
Mission Control executes, the vault endures
The resolution came from a pattern every data architect knows: you do not run analytics against the order-entry database, and you do not process transactions in the warehouse.
So the environment is now layered. Mission Control is the execution layer, the system of record for agent task execution: status, activity, cost, deliverable evidence. That is the OLTP side. The vault is the durable layer, the system of record for decisions, strategy, governance rules, and curated knowledge. That is the warehouse. The vault's work queue went back to being what it should have been all along: my human-level priority view, not an agent task tracker.
Two flows connect the layers. Rules flow down: vault policy binds all agents, and the runtime's local operating docs implement it locally while staying authoritative for runtime-internal operations. Material events flow up: decisions and run summaries get promoted to the vault as append-only files, but only for work that crosses a materiality threshold, meaning anything that touches code, clients, money, or public content, or produces artifacts expected to survive thirty days. Everything below that line stays informal, on purpose; logging everything buries the records that count.
Append-only matters too. Agents create new files rather than editing shared ones, so sync conflicts across machines are structurally avoided instead of resolved after the fact.
The decision log entry puts the whole resolution in one line: the two maps were reconciled by reference, not merger.
What a team should take from it
Any organization running AI agents alongside human governance will eventually grow two sources of truth: one where the work happens, one where the decisions live. That is not a failure. It is a signal the operational side matured faster than the paperwork.
Resist the panic merge. Name the layers explicitly, define which questions each one answers, and set the direction of flow: rules down, material events up, with a threshold so trivia never gets promoted. The test of governance is not whether you have one system. It is whether every system knows exactly which truth it owns.