Consider a scenario that has played out across early multi-agent AI deployments. Five instances of a large language model are given the same clinical case and asked to deliberate. By the third round of discussion, all five converge on the same diagnosis. It is wrong. The first agent’s confident-but-incorrect assessment anchored the others. Each round of “deliberation” did not add diversity — it destroyed it.
This is the central problem of multi-agent AI: coordination can kill the very thing that makes multiple agents valuable. Research on multi-agent conversation frameworks for clinical diagnosis, published in the Journal of Medical Internet Research, confirms that unstructured agent deliberation often produces anchoring rather than genuine analytical diversity — whereas structured frameworks that enforce independent assessment first show substantially improved accuracy.
The Anchoring Trap
Daniel Kahneman and Amos Tversky’s foundational 1974 paper on judgment under uncertainty identified anchoring bias as one of the most robust and consequential heuristics in human decision-making: once an initial estimate is formed, subsequent judgments are systematically pulled toward it, even when the anchor is arbitrary or incorrect. Decades of follow-on research in medical settings have shown that clinicians who receive a prior diagnosis before forming their own assessment are significantly more likely to agree with it — even when their independent reasoning would have reached a different conclusion.
AI agents are not immune. They may be more susceptible. Language models are trained to produce coherent, contextually appropriate text. Put three agents in a shared dialogue and ask them to deliberate, and you get a consensus factory, not a deliberation chamber. The training objective — produce the next most likely token — rewards agreement with prior context, not independent analysis.
How We Solved It: Architectural Isolation
Manthan Intelligence’s Analytical Council uses a three-phase architecture specifically designed to prevent anchoring:
Phase 1: Independent Assessment. Each analytical lens evaluates the company in complete isolation. The technology lens does not see the market lens’s output. The operations lens does not know what the financial lens concluded. They receive identical source data — the deal memo and public information — and produce independent assessments with independent confidence levels.
This is not an implementation convenience. It is the product. The entire value of multi-agent analysis collapses if agents see each other’s work before forming their own view.
Phase 2: Lock, Then Reveal. Every lens commits its assessment and confidence score before any synthesis occurs. These commitments are immutable — no lens can revise its assessment after seeing what the others said. When the synthesis layer opens, it reads 3, 6, or 9 locked assessments that represent genuine analytical diversity.
Think of it like a jury that deliberates in sealed rooms, slides their assessments under the door, and only then convenes to discuss.
Phase 3: Synthesis Under Tension. The synthesis layer does not average the assessments. It looks for productive disagreement — cases where two lenses reach opposite conclusions from the same data. A technology lens that says “strong moat” and an operations lens that says “can’t scale” are not contradicting each other; they are identifying a tension that a single analyst would have missed entirely. The synthesis layer’s job is to name that tension, not resolve it into bland consensus.
What the Knowledge Graph Does (and Doesn’t Do)
There is a shared resource: the Knowledge Graph. It contains factual data — funding rounds, market sizes, competitor lists, sector benchmarks. Every lens reads from the same KG, ensuring factual consistency.
But the KG contains zero opinions. No previous conclusions. No “last time we looked at a company like this, we said X.” The facts are shared; the judgments are independent. This is a deliberate architectural choice. A shared opinion layer would be the fastest way to create a system that agrees with itself.
The Medical Second Opinion Analogy
The best parallel is how good hospitals handle difficult diagnoses. The first specialist examines the patient and writes their assessment. The second specialist examines the same patient — same scans, same blood work, same history — but does not read the first specialist’s assessment first. Only after both have committed their independent diagnosis does the case conference begin.
When they agree: high confidence. When they disagree: the disagreement itself is the most valuable information in the room, because it identifies exactly where the clinical uncertainty lives.
Multi-agent diagnostic frameworks that enforce this kind of structured independence — with agents assigned distinct roles before any shared deliberation begins — consistently outperform both single-agent systems and unstructured multi-agent deliberation in accuracy and diagnostic value. Our Analytical Council applies the same principle, except we can run 3 to 9 independent specialists simultaneously, and every disagreement is logged as a data point that makes the next analysis better.
Why This Matters for Knowledge Workers
If you are evaluating an AI analytical tool, ask one question: do the agents see each other’s work before committing their own? If yes, you have an expensive consensus machine. If no — if the architecture enforces genuine independence — you have something that might actually catch what a single analyst would miss.
The hardest part of building a multi-agent system is not making agents talk to each other. It is preventing them from talking too early.
This analysis draws on Kahneman and Tversky’s 1974 paper on judgment under uncertainty, JMIR research on multi-agent conversation for clinical decision-making (2024), and PMC research on multi-agent diagnostic frameworks. Human editorial oversight applied.
This analysis is informational and does not constitute investment advice, a research report, or a recommendation to buy, sell, or hold any security.
Charaka Notes by Manthan Intelligence. Subscribe