Skip to main content
After both monks complete their essays, perform a critical verification step: check that the monks actually diverged. The skill’s value comes from structurally uncorrelated exploration of the problem space. Without genuine divergence, you get the same analysis with different conclusions bolted on — which produces compromise, not synthesis.

What to Check

Verify the monks actually diverged by examining:

Evidence Sources

Do the monks cite different evidence, or substantially overlapping sources?

Conceptual Vocabulary

Do they frame the problem using different conceptual vocabularies?

Unstated Assumptions

Do their unstated assumptions diverge, or do they share the same background framework?

Perspective Recognition

Would a reader recognize these as genuinely different perspectives, or the same perspective with different conclusions bolted on?

The Structural Test

The key question: Are the monks in “same framework, different conclusions” mode or “genuinely different conceptual frames” mode?
Low decorrelation looks like:
  • Both monks cite the same 5-7 sources
  • Both use the same technical vocabulary and conceptual framework
  • Both assume the same unit of analysis (e.g., both analyze at the individual level, or both at the organizational level)
  • The only difference is which conclusion they reach
High decorrelation looks like:
  • Different evidence bases
  • Different conceptual frameworks (one uses economic terminology, the other uses psychological)
  • Different units of analysis (one focuses on individual incentives, the other on systemic constraints)
  • Different metaphors and analogies
  • Reading both essays feels like encountering two genuinely different worldviews

Why Decorrelation Matters

From the ensemble diversity research (Wood et al., JMLR 2023):
The bias-variance-diversity decomposition shows diversity is literally subtracted from ensemble error: E[loss] = noise + avg_bias + avg_variance − diversity
Correlated errors eliminate the diversity benefit entirely. This is why monks must be spawned in separate sessions with no shared context, and why heterogeneous model families (when available) increase the skill’s creative output.
Surowiecki’s wisdom-of-crowds conditions: Independence is necessary, not optional. Correlated agents produce crowd madness, not wisdom.

What to Do When Decorrelation is Low

Option 1: Reformulate the Belief Burdens

Consider reformulating the belief burdens to force genuinely different conceptual frames, not just different positions within one frame. Ask yourself:
  • Are both monks operating from the same implicit model of how the domain works?
  • Are both assuming the same thing is the central question?
  • Could you assign belief burdens that force different units of analysis or different ontological claims?

Option 2: Revise the Framing Corrections

The framing corrections in the monk prompts may not have been strong enough. Review them:
  • Did you identify the degenerate framing and explicitly steer monks away from it?
  • Did you give each monk an ontological question that forces them to conceptualize the domain differently?
  • Did you warn each monk what their opponent’s actual strongest argument is (not the strawman version)?

Option 3: Use Heterogeneous Models

If available, spawn the monks using different model families:
  • Different training data produces different “intuitions”
  • Different blind spots
  • Different reasoning patterns
  • Different default framings
This is structural decorrelation at the training-data level, which is the single most promising direction in the multi-agent debate literature (Du et al., ICLR 2025).

Model Heterogeneity

When possible, use different model families for Monk A and Monk B. The orchestrator should remain your strongest available model (it needs maximum synthesis capability), but monks benefit from heterogeneity.

Option 4: Restart with Better Prompts

If decorrelation is genuinely low — the monks are arguing from the same conceptual framework — prefer restarting with revised prompts over continuing. Fresh context with better instructions produces better results than trying to salvage monks that have locked onto the same frame.

How to Strengthen Decorrelation in Prompts

1. Assign Different Ontological Claims

Force each monk to answer “What IS this thing?” differently. Example from TanStack/Next.js dialectic:
  • Monk A: “A framework is a tool for sovereignty — the proper relationship is infrastructure as open primitive”
  • Monk B: “A framework is a tool for capability — the proper relationship is deep integration with commercial infrastructure”
Both are arguing about frameworks, but they have fundamentally different concepts of what a framework is for.

2. Assign Different Units of Analysis

Example:
  • Monk A analyzes at the individual contributor level (developer experience, learning curve, day-to-day workflow)
  • Monk B analyzes at the organizational level (team coordination, maintenance burden, hiring constraints)

3. Assign Different Temporal Frames

Example:
  • Monk A reasons from historical patterns (“every time this structure emerged in the past, it led to X”)
  • Monk B reasons from future possibility (“the landscape is shifting in ways that make new approaches viable”)

4. Target Different Evidence Domains

Give each monk specific research directives that pull from different knowledge domains. Example:
  • Monk A: “Search for economic/incentive analyses of this structure”
  • Monk B: “Search for cognitive/psychological research on how people actually use these tools”

The Decorrelation-Synthesis Connection

Strong decorrelation in Phase 3 enables strong synthesis in Phase 5. When the monks genuinely inhabit different conceptual frameworks:
  • The shared assumptions they reveal (Phase 4.2) are more surprising and more powerful
  • The Boydian decomposition (Phase 4.5) finds richer cross-domain connections
  • The synthesis has to do genuine conceptual work to bridge the gap
Weak decorrelation produces weak synthesis — because there’s no real gap to bridge, just a preference to split.
Check decorrelation before investing in Phase 4-5. If the monks didn’t diverge structurally, the downstream analysis will be shallow no matter how carefully you execute it. Better to restart the monks than to push forward with correlated essays.

Build docs developers (and LLMs) love