Last week a desk research mandate landed in my inbox. European food and beverage companies preparing their entry into India, now that the EU-India trade agreement is signed. This is the kind of study a big consulting firm staffs with three Business Analysts and bills six figures. I’m running it with AI agents. And I’m testing a method I haven’t seen anyone in consulting use yet. I call it the gauntlet loop. The first move is still the classic one. You don’t start by “researching the market”. You write your gut answer on day one, then spend the rest of the study trying to kill it. My day-one hypothesis was simple: The winners of this FTA are not the biggest brands. They are the companies that sell the categories where the tariff math actually moved. Wine drops from 150% to a fraction of that. Olive oil goes to zero. Dairy is excluded entirely. That is three different games inside one “F&B opportunity”. So the scope is not an overview of India. It is a clean issue tree. Mutually exclusive buckets that cover the full question. Which categories win on tariff math. Which route to market fits each one. What the real import chain looks like. Who has already moved and what they plan next. Nothing overlapping. Nothing missing. The same discipline you would use if you were preparing a case for an MBB interview. In MBB the method has never been the problem. The problem is that this method is normally executed by a pyramid of 25-year-olds billed at partner-adjacent rates, takes six months, and often ends up too far from reality for an SMB to implement it. Most people solve this by handing the study to a research agent. I operate agent fleets every day, so I can say this plainly: An unsupervised research agent produces confident garbage at scale. Not because the models are bad. Because the agent grades its own homework. Ask it if the work is done and it says yes. Every time. It spends more time reviewing the source than challenging what the source actually claims. The real bottleneck is not generation speed. It is the quality bar, and the fact that agents drift toward the sources they found instead of staying locked on the original question. A few weeks ago a pattern started spreading in the coding-agent world: the gauntlet loop. The mechanics are simple and strict: An orchestrator that never writes the deliverable itself. Builder agents that produce. Blind critics with fresh context that judge the artifact against a concrete reference example. And the rule that makes it work: a critic never reviews a second attempt. New round, new critic. A critic who already saw draft one starts grading “better than before” instead of “meets the bar”. The loop only stops on inspectable evidence. Never when the builder declares itself done. Coders use this to ship features. I have not seen anyone run it on consulting deliverables. So that is the experiment. For this study the reference artifact is the strongest comparable market study I could find, torn down to its skeleton. One study done for another locality. Another done for the same locality but on a different sector. I looked at how the executive summary argues. How the country one-pagers are built. Where the buyer-side detail sits. That skeleton became the quality bar. One builder agent per issue-tree bucket. Each finished section goes to a blind critic that has never seen our study. Below the bar → rejected with written reasons, rebuilt, judged again by a fresh critic. A separate pass does nothing but check claims against their sources. What stays human is clear: The issue tree. The interviews (real importers do not talk to bots). The “so what” on every page. The decision to kill a hypothesis when it dies. Judgment stays human. The pyramid cost and the six-month timeline do not. This is what a mini-MBB looks like in 2026. Same scoping discipline. Agent execution. Gauntlet quality control. Partner-level judgment on top. Faster because the desk work takes hours instead of weeks. Cheaper because there is no pyramid to feed. And the same loop can run next week on a proposal, a due diligence or a competitive teardown. If you run a boutique B2B firm, the real point is this: The frameworks were never the moat. The pyramid was just the cost of the organisation. That cost is leaving the building. Some people say this is the end of consulting. I think this is how we revive it and make it usable for a much larger number of firms. I will share what the gauntlet actually catches, and what it misses, once the study ships.