Mycroft Finance Investigator: Evidence Gap Planner & Specialist Routing

Supervised teams vs single agents: Lineage verification before delegation

[0:00]Hi, I am Adwait Changmai, and this video is about turning a single finance agent into a small team of specialists. This week I built a supervisor that ranks evidence gaps by money and routes each one to the specialist who can chase it. The interesting part is what sits at the top of that queue because no specialist

Arithmetic gap ranking: Prioritizing gaps by cumulative adverse impact ($)

[0:16]In the system can close it. The supervisor does three things strictly in that order. It verifies the entire source chain first. Only then does it rank the outstanding evidence gaps by how much money they account for, and only then does it hand each one to a

The $200k revenue gap: Why top-ranked items wait on owner explanations

[0:30]Specialist. Nothing is delegated until the data underneath has been checked. Four delegations in total, and the first one is the one people skip. Before any category is analyzed, a lineage specialist is asked to verify the input chain. Only after that comes back verified do the revenue and cost specialists get any work. Notice that two categories share one specialist.

Fencing specialists: Scoped boundaries, capped delegations, and fingerprinting

[0:50]This is a team, not one agent per row. And this is the queue it produced. Revenue first at 200,000 of cumulative adverse impact, cost of goods second at 100,000, operating expenses third at 36,000. The ranking is not a judgment about importance. It is arithmetic over verified periods, which means a reviewer can recompute it. Now the part I

Humans in the loop: Why a prioritized queue is not an approved decision

[1:11]Actually want to show you. Look at what each of those gaps is waiting on. Not more computation, not a better model, something the system is structurally incapable of producing. Owner causal explanation. That is the missing evidence on the highest ranked gap in the queue worth 200,000, and it is a sentence only a human who runs the business can write. The other two want operational driver records as well, which also come from outside the system. The supervisor can rank the work. It cannot finish any of it. The ranking itself is 10 lines, and the condition on line four is the whole design. Only periods already marked materially adverse are counted. A category is not promoted because it looks dramatic, or because it appeared most often. It is promoted because verified periods added up to more money. And because specialists are agents with their own scope, each one is fenced. A specialist can only touch its assigned category. There's a hard limit on how many delegations a run

May make. Every specialist result carries a fingerprint, and any change to a source file after the fact is detected rather than absorbed. Which leaves a very clear line down the middle of this system. Specialists collect and verify evidence. They do not infer causation, and they do not approve anything. The report says so in its own header, and it ends with a sentence I put there on purpose. A prioritized queue is not an approved explanation or decision. The temptation with a specialist system is to let the top agent summarize and call that an answer. I did not build that. The supervisor produces a work queue with a named owner on every row, and the run stays blocked until a person does the part only a person can do. So, the verdict. The Mycroft finance investigator now runs as a supervised team rather than a single agent. Before anything is delegated, a lineage specialist verifies the whole source chain. Then outstanding evidence gaps are ranked by cumulative adverse

Impact. Revenue at 200,000, cost of goods at 100,000, operating expenses at 36,000, and routed across four recorded handoffs. Specialists are scoped. Delegations are capped, and every result is fingerprinted, and every ranked gap is blocked on a human because the evidence it needs is an owner's explanation that no agent in the system can write. Your turn. Here is the prompt. Take a multi-agent system I am building and tell me, for every task it delegates, what evidence would close that task, and whether any agent in the system can actually produce it. Then list the tasks that are secretly waiting on a human. Run that against something you have built. The second half is where the surprises are because a queue that looks automated is often a queue of things nobody has assigned to a person yet. That was the evidence gap planner and specialist routing for the Mycroft finance investigator. The code and the full handoff trace are in pull request 17 at wait changen for humanitarian's AI.