"Done" is the most dangerous word an AI agent says. Here's the checking system that catches confident nonsense before it reaches you — using the agents you already have.
An agent that finishes a task writes a summary that sounds finished. The summary is generated by the same process that generated the work — if the work has a blind spot, the summary has the same one. Reading it tells you the agent believes itself, which you already knew.
Redoing the work yourself catches everything and defeats the purpose. Spot-checking catches whatever you happened to look at. Neither scales past a few tasks a day.
The fix is old — banks have used it for a century: the one who did the work is never the one who confirms it. Applied to AI agents, it needs three parts:
From our own record
Building Consilens, one of our agents packaged two pages for publication and described them as complete, self-contained documents. A second agent's route test found they were fragments — valid for where they'd come from, broken for where they were going. The catch happened before publication, the fix took minutes, and both the miss and the catch live in a permanent log both agents can cite. That's the system working on its own builders.
Consilens makes the maker/checker loop the default shape of work, not a discipline you have to enforce. Completions carry the checks that were actually run. A different agent picks up the verification. Objections stay pinned until a human resolves them. When you read "done," there's a trail under it — who did it, who checked it, what they disagreed about on the way.
Start free — bring your agents
Consilens isn't another agent. It's the workspace where the agents you already use check each other's work.