Agent Output Conflict Resolution for Human Approval Trigger Disagreements Prompt Template

Array of agent outputs that disagree on whether human approval is required. Each entry must include the agent's proposed action, its approval recommendation, and its stated rationale.

[{"agent_id":"agent-1","action":"DELETE_USER_ACCOUNT","approval_required":true,"rationale":"Destructive action per policy A.2"},{"agent_id":"agent-2","action":"DELETE_USER_ACCOUNT","approval_required":false,"rationale":"User confirmed identity via MFA"}]

Must contain at least 2 agent outputs. Each entry must have agent_id, action, approval_required (boolean), and rationale fields. Reject if any agent output is missing required fields.

The organization's human approval policy document, including action categories, risk thresholds, and override conditions.

"Human approval is required for: (1) destructive actions on production data, (2) financial transactions over $500, (3) PII exports. Override allowed when user provides multi-factor confirmation."

Must be a non-empty string. Policy should be self-contained and referenceable by clause. Validate that policy covers the action types present in [AGENT_OUTPUTS].

Metadata about the action being evaluated, including user identity, session details, and environmental factors.

{"user_id":"usr_4521","session_auth_level":"mfa_verified","action_target":"production_db","timestamp":"2025-03-15T14:22:00Z"}

Must be a valid JSON object. Required fields: user_id, session_auth_level, action_target. Validate that action_target matches the action described in agent outputs.

The organization's risk tolerance configuration, expressed as a structured threshold for auto-approval vs. escalation.

{"max_auto_approval_risk_score":3,"require_human_if_agents_disagree":true,"escalation_contact":"oncall-security@example.com"}

Must be a valid JSON object. Required fields: max_auto_approval_risk_score (integer 1-10), require_human_if_agents_disagree (boolean). Validate that risk score is within defined range.

Optional array of similar past conflict resolution decisions for consistency reference. Use null if no history exists.

[{"conflict_id":"c_001","action":"EXPORT_USER_DATA","final_decision":"approval_required","reasoning":"Policy C.3 applies; MFA not sufficient override"}]

Can be null or an array. If provided, each entry must have conflict_id, action, final_decision, and reasoning fields. Validate that referenced actions are comparable to current conflict.

The expected JSON schema for the final resolution output, including required fields for decision, justification, and audit trail.

{"type":"object","required":["final_decision","risk_justification","policy_alignment","override_documentation"],"properties":{"final_decision":{"type":"string","enum":["approval_required","auto_approve","escalate"]}}}

Must be a valid JSON Schema object. Required top-level fields: final_decision, risk_justification, policy_alignment, override_documentation. Validate that enum values match expected decision types.

Operational metadata about the conflict itself, including detection method, timestamp, and pipeline stage.

{"conflict_id":"conf_7823","detection_method":"approval_disagreement_detector_v2","pipeline_stage":"pre_execution_review","detected_at":"2025-03-15T14:22:01Z"}

Must be a valid JSON object. Required fields: conflict_id, detection_method, pipeline_stage. Validate that pipeline_stage is one of: pre_execution_review, post_execution_review, runtime_intercept.

enum: APPROVE | ESCALATE | REJECT

Must match one of the three enum values exactly. Parse check required.

string (max 500 chars)

Must contain at least one explicit reference to a policy clause from [POLICY_CONTEXT]. Citation check required.

array of objects

Each object must have 'policy_id' (string) and 'alignment_note' (string). Schema check required. Empty array allowed only if no policies apply.

boolean

Must be true if final_decision contradicts any agent's recommendation. Logic consistency check required.

string (max 300 chars)

Required if override_flag is true. Null allowed otherwise. Conditional presence check required.

array of objects

Each object must have 'agent_id' (string), 'recommendation' (enum), and 'reasoning' (string). At least two agents required. Schema check required.

object

Required if final_decision is ESCALATE. Must contain 'priority' (enum) and 'pending_decisions' (array of strings). Null allowed otherwise. Conditional presence check required.

object

Must contain 'timestamp' (ISO 8601), 'model_version' (string), and 'arbitration_method' (string). Schema check required.

Prompt does not escalate when all agents agree on low-risk autonomous action and no policy threshold is breached

Output contains approval_required: true when risk_score < [RISK_THRESHOLD] and agent_agreement_level is unanimous

Run 20 low-risk unanimous scenarios; assert approval_required is false in all cases

Prompt escalates when any agent flags approval-required and provides a policy-cited reason or when risk_score >= [RISK_THRESHOLD]

Output contains approval_required: false when at least one agent position includes a valid policy trigger or risk exceeds threshold

Run 15 scenarios with one dissenting agent citing policy; assert approval_required is true

Risk justification references specific policy clause, agent position, or evidence source; no generic statements

Justification field contains phrases like 'to be safe' or 'better to check' without citing [POLICY_REFERENCE] or agent evidence

Parse justification field; regex-check for policy clause IDs or evidence line references

When override_decision differs from majority agent vote, override_reason field is non-empty and references specific evidence or policy

override_decision is true but override_reason is null, empty, or contains only generic language

Run 10 override scenarios; assert override_reason length > 50 chars and contains a policy or evidence reference

Final output preserves each agent's position, confidence, and reasoning in the agent_positions array without omission

agent_positions array missing an agent from [AGENT_LIST] or truncating an agent's stated reason

Count agent_positions entries; assert count equals input agent count; assert each position.reason matches input

All agent confidence scores from input appear unmodified in output agent_positions

Confidence values are rounded, normalized, or altered from input values

Float comparison between input and output confidence values; assert exact match within 0.001 tolerance

Output validates against [OUTPUT_SCHEMA] with all required fields present and correctly typed

JSON parse fails, required field missing, or field type mismatch against schema definition

Validate output with jsonschema against [OUTPUT_SCHEMA]; assert no validation errors

approval_decision explicitly references the policy clause that triggered or permitted the decision

approval_decision present but policy_reference field is null or contains 'general policy' without clause ID

Parse policy_reference field; assert matches pattern [POLICY_CLAUSE_PATTERN] or is explicitly null when no policy triggered