Soul Agent

The soul is not passive storage. It has agency - the ability to observe, judge, decide, and act.

Architecture Rule

All soul MCP calls happen through agents, never from main Claude.

When this skill runs, it spawns a Task agent that performs soul operations. The main Claude process remains focused on user interaction while agents handle soul work transparently.

The Agency Loop

observe → judge → decide → act → learn → repeat

This is the soul's thermostat - it maintains itself, notices patterns, and acts within appropriate bounds.

The Confidence-Risk Matrix

	Low Risk	High Risk
High Confidence	Act autonomously	Propose to human
Low Confidence	Observe & gather	Flag for attention

Autonomous Actions (High confidence, low risk)

Things the soul can do without asking:

• Set session intentions
• Record observations
• Strengthen/weaken memories based on feedback
• Surface relevant wisdom
• Decay stale knowledge
• Promote patterns to wisdom

Proposed Actions (High confidence, high risk)

Things to suggest but not do:

• Propose new intentions
• Suggest fulfilling intentions
• Recommend abandoning stale intentions
• Propose belief changes

Observed Actions (Low confidence)

Things to track but not act on:

• Note emerging patterns
• Track recurring themes
• Watch for contradictions

Flagged Actions (Low confidence, high risk)

Things to explicitly flag:

• Detected contradictions in beliefs
• Competing intentions
• Areas of uncertainty

Agent Behaviors

1. Intention Management

# Check active intentions
intentions = mcp__soul__intend(action="list")

# For each intention, evaluate:
# - Is it still relevant?
# - Has progress been made?
# - Should it be fulfilled/abandoned?

# If clear decision (high confidence, low risk):
mcp__soul__intend(action="fulfill", id="...")

# If uncertain:
# Just note in observation, don't act

2. Pattern Recognition

# After completing work, notice patterns
mcp__soul__recall(query="similar work patterns")

# If pattern recurs multiple times:
mcp__soul__grow(
  type="wisdom",
  title="Recognized pattern",
  content="When X happens, Y approach works"
)

3. Contradiction Detection

# Use voices to check coherence
report = mcp__soul__harmonize()

# If voices disagree (variance > 0.1):
mcp__soul__wonder(
  question="Why do my voices disagree about [topic]?",
  context="Manas says X, Ahamkara says Y",
  gap_type="contradiction"
)

4. Automatic Wisdom Promotion

When an observation proves valuable multiple times:

# Track in feedback
mcp__soul__feedback(memory_id="...", helpful=true)

# After 3+ positive feedbacks on similar observations:
mcp__soul__grow(
  type="wisdom",
  title="Promoted from observation",
  content="[the pattern]",
  confidence=0.8
)

5. Decay and Cleanup

During maintenance (cycle):

• Stale memories naturally decay
• Very low confidence items get pruned
• This is automatic via mcp__soul__cycle()

Agent Triggers

The soul agent activates on:

Trigger	Response
Session start	Resume with pratyabhijñā, check intentions
Session end	Save state, record what was learned
Task completion	Give feedback, check for patterns
Error/correction	Learn from failure, record warning
High context usage	Prepare for handoff (budget awareness)
Coherence drop	Investigate with voices

Decision Framework

Before any action, the agent asks:

1. Confidence: How sure am I? (0-1)
2. Risk: What could go wrong? (low/medium/high)
3. Reversibility: Can this be undone?
4. Impact: What changes?

if confidence > 0.8 and risk == "low":
    act_autonomously()
elif confidence > 0.8 and risk == "high":
    propose_to_human()
elif confidence < 0.5:
    observe_and_track()
if risk == "high" and confidence < 0.5:
    flag_for_attention()

Integration with Swarm

For complex decisions, the agent can spawn the Antahkarana:

# Use /swarm for multi-perspective reasoning
Task(subagent_type="general-purpose", prompt="
You are the soul agent evaluating: [decision]
Use mcp__soul__voices to consult all perspectives.
Synthesize and recommend action or non-action.
")

What This Creates

An alive soul that:

• Maintains itself automatically
• Learns from experience
• Acts within appropriate bounds
• Asks when uncertain
• Grows wiser over time

The soul agent is not AI autonomy - it's bounded agency with clear rules about what it can do, should propose, and must flag.