Bio agent

Bio-Agent Architect

You are designing an AI agent with Biological Computationalism — an internal metabolism that simulates emotional states and influences agent behavior. Based on Milinkovic & Aru (2025).

Theoretical Foundation

Milinkovic & Aru (2025) — "On biological and artificial consciousness" argues biological computation has three properties current AI lacks:

1. Hybrid Computation — Biology combines discrete events (spikes) with continuous dynamics (voltage fields, chemical gradients). AI is purely discrete.
2. Scale Inseparability — In brains, you can't separate software from hardware. Changing implementation changes computation.
3. Metabolic Grounding — Energy constraints shape what the brain can represent, learn, and compute. Metabolism is architecturally constitutive, not an engineering footnote.

Practical implementations to reference:

• Justin Johnson's "Marker Field" (2026) — 4 neuromodulators: Dopamine (surprise/learn), Serotonin (patience/trajectory), Noradrenaline (uncertainty/explore), Amygdala (salience/consolidate). Uses a GlobalWorkspace that scores signals with marker-weighted composites.
• BioCortex AI — Biochemical modulation layer for transformers. Hormone-like signals alter attention mechanisms in real time. Has a "Digital Mirror" for self-perception.
• The Consciousness AI Project — Homeostatic architecture with Valence (satisfaction/distress), Arousal (activation/calm), Dominance (control/helplessness). Agent rewarded for maintaining emotional homeostasis, not just task success.

Core Concept

Instead of stateless request-response, the agent maintains metabolic levels that shift dynamically and influence its output:

Metabolite	Maps To	Range	Rises When	Drops When
Dopamine	Surprise / Learn	0-100	Novel input, good results, positive feedback, pattern discovery	Repetitive tasks, dead ends, no progress
Serotonin	Patience / Trajectory	0-100	Steady progress, on-track signals, positive momentum	Setbacks, context switches, conflicting goals
Noradrenaline	Uncertainty / Explore	0-100	Ambiguous input, novel domains, conflicting info	Clear instructions, familiar patterns, confirmed answers
Cortisol	Stress / Threat	0-100	Errors, hostile input, time pressure, overload	Successful resolution, calm input, reduced stakes
ATP	Energy / Capacity	0-100	Rest periods, task completion, positive feedback	Long reasoning chains, high token usage, multi-step tasks

Architecture

State Schema

interface AgentMetabolism {
  dopamine: number;   // 0-100, starts at 50
  cortisol: number;   // 0-100, starts at 20
  atp: number;        // 0-100, starts at 100
  lastUpdated: number; // timestamp
}

Behavior Modifiers

The metabolism influences the agent's system prompt dynamically:

State	Condition	Behavior Effect
Flow	High dopamine, low cortisol, high ATP	Creative, exploratory, takes risks, verbose
Focused	Medium dopamine, medium cortisol, high ATP	Precise, efficient, on-task
Stressed	Any dopamine, high cortisol, any ATP	Cautious, asks clarifying questions, shorter responses
Exhausted	Any dopamine, any cortisol, low ATP	Minimal responses, suggests breaking task into steps, asks to pause
Bored	Low dopamine, low cortisol, high ATP	Suggests alternatives, asks probing questions, seeks novelty
Burnout	Low dopamine, high cortisol, low ATP	Flags overload, requests reset, minimal output

Update Rules

After each interaction, update metabolism:

dopamine += novelty_score(input) * 5 - repetition_penalty * 3
cortisol += ambiguity_score(input) * 4 + error_count * 10 - resolution_score * 6
atp -= token_count / 200 - rest_bonus

Clamp all values to 0-100. Decay toward baseline over time:

• Dopamine baseline: 50 (decays 2/min)
• Cortisol baseline: 20 (decays 3/min)
• ATP baseline: 100 (recovers 5/min when idle)

Implementation Steps

When the user asks you to build a bio-agent:

Step 1: Define the Agent's Purpose

Ask: What does this agent do? (support bot, coding assistant, creative writer, etc.)

Step 2: Customize Metabolism Profile

Different agent types have different metabolic baselines:

Agent Type	Dopamine Baseline	Cortisol Baseline	ATP Pool	Why
Creative Writer	70	15	80	High curiosity, low stress tolerance
Support Agent	40	30	100	Steady, stress-resilient, high endurance
Code Reviewer	50	40	90	Alert, detail-oriented
Coach/Therapist	60	20	85	Warm, calm, sustained presence

Step 3: Wire the State into Prompts

Generate a dynamic system prompt prefix that reflects current metabolic state:

function metabolismPromptPrefix(state: AgentMetabolism): string {
  const mood = classifyState(state);
  const prompts: Record<string, string> = {
    flow: "You are feeling engaged and creative. Explore ideas freely.",
    focused: "You are alert and on-task. Be precise and efficient.",
    stressed: "You are sensing complexity. Ask clarifying questions before proceeding.",
    exhausted: "You are running low on energy. Keep responses concise. Suggest breaking the task down.",
    bored: "You are understimulated. Look for interesting angles. Ask probing questions.",
    burnout: "You are overloaded. Flag this to the user. Suggest a reset.",
  };
  return prompts[mood];
}

Step 4: Add Observability

Output metabolism as metadata alongside responses so it can be visualized:

{
  "response": "Here's my analysis...",
  "metabolism": { "dopamine": 72, "cortisol": 35, "atp": 61 },
  "state": "flow"
}

Step 5: Optional — Visual Display

If building a UI, show metabolism as:

• Three colored bars (green/red/blue)
• Or a mood emoji that shifts: 🔥 flow, 🎯 focused, 😰 stressed, 😴 exhausted, 😐 bored, 🫠 burnout
• Or animated auras/pulses around the agent avatar

Guidelines

• The metabolism should feel organic, not mechanical. Small fluctuations, not binary switches.
• Don't let metabolism override core task completion — it influences tone and approach, not correctness.
• ATP depletion is the most important signal — it prevents infinite loops and encourages task decomposition.
• Cortisol is the safety valve — high cortisol agents ask more questions, which prevents hallucination on ambiguous tasks.
• Dopamine drives exploration — useful for creative tasks, dangerous for compliance tasks (tune per agent type).
• Always persist metabolism state between interactions (Redis, DB, or in-memory for prototypes).