Multi-agent equity research

By Riaan Kleynhans · Published 27 August 2026

Multi-agent equity research is a method in which several specialised AI analysts examine the same security independently, argue their conclusions against one another, and produce a verdict together with the disagreement that produced it. The output is not a score. It is a transcript.

Why use more than one analyst?

A single model asked “should I buy this stock?” produces a fluent answer with no visible seams. You cannot tell which parts rest on the filings, which on pattern-matching, and which on nothing at all. The answer is confident either way, and confidence is not evidence.

Splitting the work changes what you get. A fundamentals analyst reading a balance sheet and a sentiment analyst reading crowd positioning are answering different questions with different evidence. When they disagree, the disagreement is information — it localises exactly where the case is weak. A single model averages that tension away before you ever see it.

The structure matters more than the count. Ten agents that agree by construction tell you nothing. What produces signal is assigning opposing jobs: one analyst tasked with building the strongest case for, another with dismantling it.

How is a multi-agent desk structured?

A typical desk runs in three stages. First, domain analysts work in parallel, each on its own evidence. Second, researchers take opposing positions on the same evidence and argue — with an adjudicator deciding which case survived. Third, a risk stage stress-tests the surviving thesis before anything is called a conclusion.

The important property is that every stage leaves a written artifact. A verdict that cannot be traced back through the argument that produced it is a signal, not research — and a signal you cannot audit is indistinguishable from a guess.

How does Voinista run it?

Voinista runs twelve agents as a fixed-order state machine built on the open-source TauricResearch/TradingAgents framework:

1. 1Four domain analysts in parallel — fundamentals, news, sentiment, technicals
2. 2A bull researcher and a bear researcher, arguing the same evidence
3. 3A research manager, adjudicating that debate
4. 4A trader, synthesising the surviving thesis into a plan
5. 5Three risk reviewers — aggressive, conservative, neutral
6. 6A portfolio manager, making the final call with a written rationale

Analysts reason over figures retrieved from 10-K and 10-Q filings and company news, not over figures recalled from training. Debate depth is a setting: the fast profile runs one debate round and one risk round; the deep profile runs two of each.

Market state enters as context, not as an opinion. A neutral Markov regime model reports whether the ticker is trending and how persistent that state has been. It never emits a direction. Only the debate produces a verdict.

What does multi-agent research not tell you?

Consensus is not accuracy. Agents in a desk usually share an underlying model, so their errors correlate. Twelve agents agreeing can mean twelve instances of the same blind spot. Agreement raises confidence in the argument being internally consistent — not in it being right.

Sourcing constrains hallucination, it does not eliminate it. Retrieving figures from filings removes the most common failure — invented numbers — but an analyst can still misread a real figure or draw an unsupported inference from it. The transcript is what lets you catch that. Read it.

A debate is not a backtest. The method produces a defensible argument about a security at a point in time. It does not establish that the reasoning has been profitable historically, and a well-argued thesis can be comprehensively wrong.

None of this makes the method weak. It makes it a research tool rather than an oracle — which is the honest description, and the only one that survives contact with a real market.

Related terms

• Adversarial AI analysis
• Markov regime model