What If AI Agents Disagreed Instead of Always Reaching Consensus?

I've been exploring multi-agent AI architectures and kept coming back to one question:

What if AI agents were designed to challenge each other rather than simply agree with each other?

This question led me to build Konsil, an open-source adversarial AI medical case-review board for Claude Code.

GitHub: https://github.com/arbade/konsil

The problem with multi-agent consensus

Multi-agent systems are becoming increasingly popular. We assign different roles to different agents, let them discuss a problem, and expect the combined result to be better than a single model's output.

But there's an interesting problem: more agents don't necessarily mean better reasoning.

Agents can converge too quickly, reinforce shared assumptions, overlook alternative explanations, and produce an answer that sounds more confident than the available evidence justifies.

I wanted to experiment with a different architecture.

Instead of optimizing for agreement, I wanted to preserve disagreement.

How Konsil works

Konsil organizes independent AI specialist agents into a structured review process.

The architecture has several distinct stages:

1. Blind specialist rounds

Each specialist receives the case independently, without knowing what other agents are thinking.

This is important because I wanted to reduce the influence of early consensus and shared framing.

2. The Challenger

After the initial assessments, a dedicated Challenger agent reviews the emerging positions.

Its job is to question assumptions, explore alternative explanations, identify missing evidence, and challenge premature conclusions.

It must also acknowledge when there is no viable challenge rather than manufacture disagreement.

3. The Checklist Officer

A dedicated agent checks whether important findings have been addressed, ruled out with evidence, or silently ignored.

The goal is to catch omissions that might otherwise disappear during the discussion.

4. The Steward

Not every possible next step is necessarily useful.

The Steward focuses on cost-aware sequencing, less invasive options, and the value of additional information.

5. The Dissent Ledger

This is one of my favorite design decisions.

Disagreements are not simply merged into a single conclusion.

They are preserved in the final board minutes, including the reasoning and confidence of individual specialists.

6. The Reviewer Gate

Before the final output is released, four independent auditors review it:

  • Citation integrity
  • Dissent integrity
  • Safety language
  • Completeness and traceability

A failed review requires corrections and another audit.

Evidence should be earned, not generated

Another important design principle is citation verification.

Agents must retrieve and verify the sources they rely on. Search queries and retrieval logs are retained so that claims can be traced back to their sources.

Unverified claims are explicitly marked instead of being presented as established facts.

A deterministic medical imaging pipeline

Konsil also includes a separate DICOM-to-3D visualization pipeline.

It uses tools such as:

  • dcm2niix
  • TotalSegmentator
  • Marching Cubes
  • NiiVue

The pipeline generates interactive 3D scenes, segmentation overlays, and multiplanar reconstruction views.

One important architectural boundary: the standard board reasoning uses written radiology reports and structured imaging outputs, not AI interpretations of image pixels.

An explicitly consent-gated experimental observational image-reading workflow is separate from the formal board process and cannot replace an official radiology report.

Testing the architecture against its own failures

One of the most interesting parts of building Konsil was discovering that the review mechanisms could catch problems generated by the system itself.

During blinded development runs, the Reviewer Gate caught:

  • Hallucinated author names in otherwise valid references.
  • A disagreement that had been silently removed from the final minutes.
  • A composite quotation incorrectly presented as verbatim.

These findings reinforced an important idea for me:

A system designed to verify AI output must be capable of questioning its own correction process, too.

Why I open-sourced it

I wanted to build something that explores a broader question in agentic AI:

Can we design multi-agent systems where uncertainty, disagreement, evidence, and accountability are first-class architectural concepts?

Konsil is my experiment in that direction.

It's packaged as a Claude Code plugin, with readable Markdown-based agent charters, commands, skills, and templates.

Installation:

claude plugin marketplace add arbade/konsil claude plugin install konsil@konsil-marketplace -y

GitHub repository:
https://github.com/arbade/konsil

I'd love your feedback

I'm particularly interested in hearing from developers working on:

  • Multi-agent orchestration
  • Adversarial AI systems
  • AI safety and evaluation
  • Evidence-grounded agent workflows
  • Claude Code plugin development

How would you approach preventing premature consensus in a multi-agent system?

What would you change in this architecture?

Konsil is experimental software, not a medical device, diagnostic tool, or substitute for qualified clinical judgment.