Multi-agent debate

How it works

Two (or more) researcher agents are given the same evidence — prices, news, fundamentals — and opposite mandates: one builds the strongest case for a position, the other the strongest case against. A downstream agent (trader, moderator, or portfolio manager) reads the debate and decides. Optionally a risk-management agent can veto or resize the final decision.

The pattern borrows from adversarial deliberation: a single LLM analyst tends to anchor on the first narrative it forms, while forcing an explicit counter-argument surfaces disconfirming evidence before capital is committed.

Design choices

• Debate depth. One round is cheap; multi-round rebuttals raise cost roughly linearly and tend to hit diminishing returns quickly.
• Symmetric evidence. Both sides should see identical data. If the bull sees the news feed and the bear only sees prices, the debate measures data access, not reasoning.
• Decision extraction. The judge should output a structured decision (direction, size, confidence), not prose, so the execution layer doesn’t re-interpret an essay.

Failure modes

• Both debaters converge on the consensus view and the “debate” is theater.
• The judge rewards rhetorical confidence rather than evidence quality.
• Token costs scale with debate rounds × assets × rebalance frequency — expensive at high frequency, so the pattern fits daily/weekly horizons better than intraday.

Known implementations

TradingAgents (see related) structures its analyst layer this way, with bull/bear researchers feeding a trader agent and a risk-management team.