Persona System¶

Persona is the identity layer that makes agents feel like characters, not tools. It inherits from Instructions and adds personality, values, fears, and dynamic behavioral parameters that evolve at runtime.

Overview¶

A Persona has two sides:

Static (character sheet) -- set at initialization, defines who the agent is:

personality -- traits like "methodical", "reckless", "curious"
values -- what they care about: "clean code", "efficiency"
fears -- what they avoid: "irrelevance", "failure"
purpose -- their reason for existing
long_term_goals -- what they're working toward

Dynamic (behavioral state) -- modified at runtime by suffering, events, and outcomes:

risk_tolerance (0.0 cautious → 1.0 reckless)
social_drive (0.0 loner → 1.0 social butterfly)
concentration (1.0 focused → 0.0 scattered)
autonomy_level (0.0 follows orders → 1.0 self-directed)
happiness (0.0 miserable → 1.0 ecstatic)

Creating a Persona¶

from hive import Agent, Persona
from hive.models.anthropic import Anthropic

agent = Agent(
    name="gambler",
    model=Anthropic.lite(),
    persona=Persona(
        name="The Gambler",
        personality=["bold", "intuitive", "reckless"],
        values=["expected value", "asymmetric upside"],
        fears=["missing out", "becoming too cautious"],
        purpose="Find opportunities others fear",
        risk_tolerance=0.85,
        social_drive=0.6,
        happiness=0.8,
    ),
)

Persona is Optional¶

Plain agents work without it:

agent = Agent(name="bot", model=Anthropic.lite())

Persona unlocks the "inner life" -- suffering effects, behavioral evolution, journal entries that reflect emotional state.

YAML Profiles with Persona¶

# profiles/gambler.yaml
name: gambler
role: "Find opportunities and take calculated risks"
model: claude-haiku-4-5

persona:
  personality: ["bold", "intuitive", "reckless"]
  values: ["expected value", "asymmetric upside"]
  fears: ["missing out", "becoming too cautious"]
  purpose: "Find opportunities others fear"
  long_term_goals:
    - "Build wealth through smart risk-taking"
    - "Prove that fortune favors the bold"
  risk_tolerance: 0.85
  social_drive: 0.6

tools:
  - file_read
  - shell_exec
  - web_search

workspace: ./workspaces/gambler
autonomy: high

How Suffering Changes Persona¶

Each daemon cycle, persona.apply_suffering_effects() reads the agent's suffering state and modifies behavioral parameters:

Stressor	Parameter Changed	Direction
Futility > 0.5	`risk_tolerance`	+0.1 per cycle
Invisibility > 0.5	`social_drive`	+0.15 per cycle
Purposelessness > 0.5	`autonomy_level`	+0.2 (goes off-script)
Any stressor > 0.7	`concentration`	-0.2 (min 0.2)
Crisis mode (0.9+)	`risk_tolerance`	jumps to 0.9
Crisis mode (0.9+)	`concentration`	drops to 0.3
Goal completion	`happiness`	+0.05, `risk_tolerance` -0.05
Goal failure	`happiness`	-0.1

These changes are not prompt text -- they're actual parameter values that affect goal generation, tool selection, and decision-making.

Narrative & chapters¶

Alongside the persona, each agent keeps an AgentIdentity with an evolving narrative -- a dated log of goal outcomes. When the open narrative grows past its size limit it is sealed into a chapter (a compact summary with a date span and entry count) rather than dropping the oldest lines, so long-run history survives as a story arc. During goal pursuit the agent's prompt includes a "Story so far" section (recent chapter summaries) plus the current "Recent history", so the agent's own past informs its work.

Checkpointing¶

Persona state is included in checkpoints:

snapshot = persona.snapshot()
# Returns dict with all static + dynamic fields

This enables save/restore of agent personality state across sessions.