Don't Trust the First Answer: The Adversarial Sub-Agent Pattern for High-Stakes Decisions

A single agent inherits your framing. You hand it a hunch, phrased as a hunch — "I think X is true — check it for me." The model reads that framing, treats your belief as the thing to support, and goes looking for evidence that fits. It comes back agreeing with you, citations included, and you feel smart. That's not research — that's confirmation bias with a tool-use loop bolted on. The fix isn't a better single prompt. It's structural: stop asking one agent to evaluate a claim, and instead run several agents that gather evidence independently, then make them argue. The answer you keep is the one still standing after the fight — not the first thing the model said to make you happy.

Modern frontier models can genuinely fan a job out to sub-agents that work in parallel — this isn't a metaphor for 'think harder.' Anthropic's own multi-agent research system uses an orchestrator that spins up several sub-agents at once, each in its own context window, each running its own searches independently before reporting back. The adversarial pattern adds two deliberate moves on top of that raw capability: you assign each sub-agent a different evidence source so they can't all drift toward the same answer, and then you add an explicit debate stage where they have to challenge each other before the orchestrator writes a single line of conclusion. Getting each agent to gather independently stops the whole answer from converging on one source. The debate step is what surfaces the counter-case you'd actually miss — the evidence you wouldn't have found if you'd asked one agent to confirm you.

This is the reusable skeleton. Swap the bracketed parts for your decision and keep the structure — the structure is what does the work. Notice the order: the hunch is stated plainly, then immediately reframed as something to attack; the sources are named and split; the debate is mandatory and comes BEFORE the report; and the model is told in plain language that you don't want the first answer, you want the one that survives.

An adversarial debate is worthless if both sides are arguing from the model's stale memory. The agents need to reach live, real data — and the moment they try to scrape it, a lot of useful sources block bots. The fix is to give the agent a real data path via a connector (MCP), so 'go check the source' actually fetches the source instead of guessing. This is generic plumbing: a scraping/fetch connector, a database connector, whatever your decision needs. The point is that each independent sub-agent has a genuine, non-blocked way to gather its own evidence.

Don't let the model bury the verdict in prose. Ask for a fixed structure so you can act fast and audit the reasoning. The shape below is the deliverable to demand at the end of the scaffold — and the best part is that a frontier model is strong enough at building HTML to render this as an interactive one-pager you can actually click through, not just a wall of text.

Strip out the example and the skeleton fits any decision where being wrong is expensive and you're tempted to trust your gut. The three sub-agents just become three different evidence angles on whatever you're deciding.