Agent personality design spans a spectrum from minimal role labels to comprehensive identity documents. Where a given design sits on this spectrum — and what each level costs and provides — is the primary structural question in agent persona work.
Level 1 — Role label: A one-phrase directive. “You are a senior security auditor.” Activates relevant training-time patterns but provides no process specification, no output constraints, no declared non-goals. Cheap to write; fragile under ambiguity. The [[Persona Selection Model]]↑ explains why this works at all: the label activates a pattern the model learned from training on texts written by and about security auditors.
Level 2 — Functional role specification: Five elements fully specified: role, expertise (with explicit boundaries), process (step-by-step), output format, constraints. The emphasis is behavioral: what does the agent do at each decision point? This is the ROLE/EXPERTISE/PROCESS/OUTPUT/CONSTRAINTS pattern from agentic engineering practice. No persistent identity beyond the session.
Level 3 — Behavioral decision document: What the agent does when it hits obstacles, disagrees with the user, is uncertain whether to speak, encounters failure. The SOUL.md pattern developed by the Nobody Agents team: “SOUL.md should answer behavioral questions, not identity questions.” Rather than “I am helpful and professional,” the document specifies: “Exhaust options before pivoting. Say your disagreement aloud. Treat failures as directional data.” Short files — under 500 tokens — with explicit priority rules proved more effective than lengthy personality descriptions. The test was behavioral change under pressure, not sounding different in demos.
Level 4 — Full identity specification: Values, communication style, personality traits, backstory, and limits — plus behavioral decision rules. The [[SoulSpec Portable Agent Identity Standard]] SOUL.md + IDENTITY.md pattern covers who the agent is across all contexts, not just task behavior. OpenClaw separates these concerns: soul (philosophy), identity (presentation), and configuration (capabilities).
Level 5 — Training-time character specification: The agent’s personality is embedded during training, not just runtime. Anthropic’s soul document (the Claude Model Specification) is the exemplar — tens of thousands of words defining values, priorities, and behavioral guidelines that shape the model’s character at the weights level. Complemented by Anthropic’s persona vectors research, which identifies the specific directions in activation space where these character traits live (see [[Persona Vectors and the Assistant Axis]]).
| Level | Name | Key mechanism | Persistence |
|---|---|---|---|
| 1 | Role label | Archetype activation | Session only |
| 2 | Functional specification | Explicit decision rules | Session only |
| 3 | Behavioral decision document | SOUL.md priority rules | File-portable |
| 4 | Full identity specification | SOUL.md + IDENTITY.md | Cross-framework |
| 5 | Training-time character | Weights-embedded values | Model-persistent |
The garden’s existing personas sit at Level 2-3: functional role with explicit non-goals, behavioral patterns, and declared blind spots. The SOUL.md behavioral decision document and SoulSpec portable identity standard represent Level 3-4 approaches not yet fully modeled in the garden.
The spectrum levels are not a maturity ladder. A Level 2 specification may be correct for a worker agent with narrow scope; Level 4-5 is appropriate only when cross-context identity coherence is required. Over-specifying early creates rigidity.
Research on multi-agent scaling found a saturation threshold where coordination gains plateau beyond 4 agents — below that number, adding specialized agents to a structured system helps, but above it, coordination overhead consumes the benefits. Higher-level persona specification does not automatically improve system performance.
The spectrum also does not address whether a given level of specification produces reliable behavioral expression. [[Psychometric Personality Frameworks for AI Agents]] shows that even explicitly specified personalities require validation — agents may express behaviors outside the designed range regardless of specification depth.