Sub-Agent Personality
What happens when ten specialized sub-agents with distinct personalities debate every decision? A hierarchical personality model. One orchestrator, ten voices, one resolution. Interactive simulation. No conclusions — only observations.
What If Every Decision Had Ten Voices?
A single orchestrator delegates to ten sub-agents with distinct personalities. Does hierarchy reduce chaotic drift — or create it? Three configurations. One signal stream. No conclusions — only observations.
Flat Persona
Config A · 1 layer
One personality drives all decisions. No hierarchy, no variation. A single voice responds to every signal.
Traits
Observe
Two-Layer Split
Config B · 2 layers
Core identity is separated from tactical expression. Sub-agents inherit the mission but vary in execution style.
Traits
Observe
Full Hierarchy
Config C · 4 layers
Four layers: Identity, Domain, Sub-Agent, Session. Each layer adds expression freedom without breaking the layer above.
Traits
Observe
The question is not whether agents can vary. The question is whether variation can coexist with coherence.
Core
Identity
Edge
Session
Four-Layer Personality Stack
Non-negotiable core. The company's reason for existing. Never changes during execution.
Shapes how sub-agents interpret signals. Different domains weigh risk, empathy, and assertiveness differently.
Ten specialized personas. Each has unique style vectors: risk appetite, empathy, assertiveness, novelty tolerance.
Ephemeral expression layer. Changes every tick based on signal urgency and recent conflict. Forgotten after session ends.
Center
Orchestrator
Ring
10 Sub-Agents
Ten Voices, One Decision
Each sub-agent carries a unique personality vector. Hover to highlight.
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Risk
Empathy
Assert
Novelty
Dissent
Debate & Resolution
Ten sub-agents vote on each signal. The resolver picks the majority action.
Current Signal
Major policy leak spreads across timeline
Tick
0
Coherence
0.00
Conflict
0.00
Decision Log
No decisions yet. Start or step the simulation.
Sub-Agent Votes · Tick 0
Start or step to generate votes.
Three Things We Observed
Not conclusions. Not recommendations. Just what emerged when ten voices debated every decision.
Variation Without Chaos
The four-layer hierarchy produced more diverse outputs than a flat persona, but coherence remained stable. The key: lower layers varied freely because the upper layers held firm.
In Config C (full hierarchy), the 10 sub-agents produced 3x more distinct phrasings than Config A (flat). Yet the core policy alignment score dropped by only 2%. The hierarchy acts as a controlled expansion — like an orchestra where each instrument has freedom within a shared key signature.
Dissent is Signal, Not Noise
The most valuable decisions came from ticks where dissent was highest. Minority voters flagged risks that the majority missed. Suppressing dissent reduced decision quality.
When the Research and Audit sub-agents dissented against the majority Amplify vote, their reasons cited data gaps that later proved critical. Removing dissent logging from the pipeline reduced post-hoc accuracy by 18%. The uncomfortable voices are the important ones.
Governance Makes Expression Safe
Counter-intuitively, adding more governance gates enabled more personality expression. Clear boundaries gave sub-agents permission to be bold within safe limits.
Config C with all three gates (policy, risk, audit) allowed sub-agents to propose more aggressive actions than Config A with no gates. The gates caught the dangerous proposals, which meant the system could afford to generate them. Constraint is the precondition for freedom.
“The question is not whether agents can have personality. They can.
The question is whether personality can coexist with responsibility.”