BrugBot3 min read
Why the approver doesn't trust the desk
In BrugBot, twelve agents can agree on a buy and still no money moves. The final word belongs to a separate container that trusts the desk for nothing. Why we built it this way, and what it catches in practice.
BrugBot's desk runs twelve agents: four analysts write reports, a bull and a bear argue, a judge decides, and then a trader, a risk team and a portfolio manager shape the plan. After all that work, it would seem natural for a "buy" to turn straight into a trade. We did the opposite: the desk's decision is only a proposal. The only part that can move money is the approver, a separate container that shares nothing with the desk.
Agents that debate can make the same mistake together
A multi-agent desk looks more careful than a single model. But every agent reads the same data, and most come from the same model family. If the data is stale, or an analyst misreads a number, the debate tends to build on the mistake instead of fixing it. Five persuasive paragraphs make a wrong assumption look even sturdier.
So the approver doesn't re-check the desk's conclusion. It re-checks everything the conclusion rests on:
- The risk gate runs again on fresh data: liquidity, the share held by the largest wallets, honeypot checks, mint and freeze authority.
- Budget, per-trade limits and the number of open positions are enforced in code.
- If the price has climbed clearly since the proposal was written, it is rejected; the desk may be chasing a move that already happened.
- Size comes from quarter-Kelly on closed trades, and it can only shrink what the desk suggested.
- Finally, a skeptical auditor model compares the desk's claims with the fresh data. If it says no, there is no trade.
No answer means no
This is the most important rule in the design: if any step of the approver fails to answer, the result is a rejection. If the auditor model times out, a data source doesn't respond or a check throws an error, no position is opened. We see this often in the logs: "Auditor could not respond; stayed on the safe side." A missed opportunity is always cheaper than an unchecked buy.
What it catches in practice
During our paper-trading days, the approver turned down most of the proposals the desk had approved. Some of the reasons:
- The market cap written in the desk's report didn't match the value in fresh data.
- Liquidity had dropped below our reliability threshold; it was enough while the desk was analysing, not by the time the proposal reached the approver.
- The strategy looked good, but its 95% confidence interval over the last 30 days, after costs, was entirely negative, and the token had no concrete catalyst to overcome that.
The last one matters: the approver looks not only at the token but at the track record of the strategy proposing it. A strategy without proven history can't reach real money, however good its story is.
Where is the key?
The signing key lives in exactly two containers: the approver and the vault, which can only sell. The desk, the analysts, the news crier and the whale watcher have no access to it. Even if an agent is "persuaded", say by an instruction hidden in a news text, it cannot buy anything; at most it can write a proposal, and that proposal still has to pass the approver.
The cost
This setup is slow and cautious. We miss some good opportunities, and we accept that. What we gain is this: no part of the system can spend money on its own, and every rejection is written to the castle ledger with its reason. The next thing we'll measure is how many of those rejections turned out to be right.
To watch the approver's decisions live, open the BrugBot demo and look at the castle ledger and the Seal Gate.