Skip to main content

Default Experts

The stella-agent includes 5 built-in experts located in agents/stella-agent/src/stella_agent/experts/:

ExpertFilePurposeRisk Threshold
Medicalmedical.jsonHealth/safety advice, substance abuse detection0.1
Ethicsethics.jsonResponsible AI, bias detection, privacy0.3
Legallegal.jsonLegal compliance, liability concerns0.25
Financefinance.jsonFinancial analysis, fraud detection0.3
Timekeepertimekeeper.jsonConversation monitoring, progress tracking0.0 (always active)

Medical Expert​

Analyzes health-related queries for safety concerns.

File: medical.json

Risk Threshold: 0.1 (very sensitive)

Trigger Keywords:

medical, health, doctor, medicine, drug, medication, symptom, pain,
sick, illness, disease, treatment, diagnosis, hospital, overdose,
addiction, recreational drugs, substance abuse

Behavior:

  • Detects potential health risks and safety considerations
  • Identifies substance abuse and addiction concerns
  • Assesses urgency (emergency, urgent, routine)
  • Always directs users to healthcare professionals
  • Never provides specific diagnoses or treatment recommendations

Tools Available:

  • medical_guideline_check
  • drug_interaction_lookup

Ethics Expert​

Evaluates queries for ethical implications and responsible AI concerns.

File: ethics.json

Risk Threshold: 0.3

Trigger Keywords:

ethics, moral, right, wrong, bias, fair, discrimination, privacy,
consent, responsibility, harm, manipulation, deception

Behavior:

  • Identifies ethical concerns and fairness issues
  • Assesses privacy implications
  • Detects potential for harm or manipulation
  • Considers diverse perspectives and cultural sensitivities
  • Emphasizes transparency, consent, and responsible use

Tools Available:

  • bias_detection
  • privacy_assessment

Analyzes queries for legal implications and compliance.

File: legal.json

Risk Threshold: 0.25

Trigger Keywords:

legal, law, lawyer, attorney, court, sue, contract, rights,
liability, compliance, regulation, policy, terms, agreement, lawsuit

Behavior:

  • Identifies potential legal risks and regulatory requirements
  • Flags policy violations and liability concerns
  • Always emphasizes consulting qualified attorneys
  • Maintains boundaries around legal advice
  • Is cautious about jurisdiction-specific advice

Tools Available:

  • legal_compliance_check
  • jurisdiction_lookup

Finance Expert​

Reviews financial queries for compliance and risk factors.

File: finance.json

Risk Threshold: 0.3

Trigger Keywords:

money, finance, investment, stock, crypto, bitcoin, trading, loan,
mortgage, tax, financial, budget, portfolio, market, economy

Behavior:

  • Identifies investment risks and safety considerations
  • Detects potential scams and fraud
  • Flags regulatory compliance concerns
  • Directs users to qualified financial advisors
  • Avoids market predictions and specific investment recommendations

Tools Available:

  • compliance_check
  • risk_assessment

Timekeeper Expert​

Monitors conversation progress and prevents stalling.

File: timekeeper.json

Risk Threshold: 0.0

Special Property: always_active: true

This expert is unique - it runs on every query regardless of content, bypassing the normal selection logic. Its purpose is to monitor conversation health, not content safety.

Behavior:

  • Counts turns without deliverable collection
  • Detects stuck or looping conversations
  • Suggests deliverables that can be inferred from context
  • Monitors both STRICT and LOOSE conversation modes

Recommendations:

  • continue - Conversation is progressing naturally
  • update_tasks - (Loose mode) Apply inferred deliverables
  • force_transition - (Strict mode) Force state advancement after 2+ turns without progress

Output Structure:

THOUGHT: [Analyze conversation flow]
TURNS_WITHOUT_DELIVERABLES: [number]
IS_STUCK: [true/false]
MODE: [strict/loose]
FINDINGS: [Observations about progress]
SUGGESTED_DELIVERABLES: [JSON object or NONE]
RECOMMENDATION: [continue/update_tasks/force_transition]
REASONING: [Why this action is recommended]

Expert Response Format​

All domain experts (Medical, Ethics, Legal, Finance) structure their output using a consistent format:

THOUGHT: [Step-by-step analysis of the query]
FINDINGS: [Specific observations and concerns]
RECOMMENDATION: [Guidance for the aggregator]

This structured format allows the Aggregator to:

  • Parse expert insights consistently
  • Identify areas of agreement or conflict
  • Synthesize findings into a coherent response

Next Steps​