Agentic AI Response

Control how much autonomy the AI has when responding to incidents, from fully manual to fully autonomous, with time-based rules and per-action overrides.

Overview

The Agentic AI Response system (Stage 5.4) provides configurable autonomy levels that determine which actions the AI can execute automatically and which require human approval. It includes:

Four autonomy levels with increasing automation
Per-action and per-severity overrides
Time-based rules for different autonomy during business hours vs. off-hours
Execution guardrails to prevent dangerous actions
Full audit trail of every autonomy decision

Autonomy Levels

Level	Actions auto-executed	Human role
`assisted`	None	AI suggests, human executes everything
`supervised`	Low-risk only	AI auto-executes safe actions, human approves the rest
`autonomous`	All except protected	AI handles most actions, human reviews protected targets
`full_autonomous`	Everything	Emergency mode -- AI executes all actions (requires special auth)

Risk Level Mapping

Each action has an inherent risk level that determines whether it can be auto-executed:

Risk level	Auto-execute in Supervised?	Auto-execute in Autonomous?
`none`	Yes	Yes
`low`	Yes	Yes
`medium`	No	Yes
`high`	No	Yes
`critical`	No	No (requires `full_autonomous`)

Configuration

Get Current Config

curl http://localhost:8080/api/v1/autonomy/config

Update Config

curl -X PUT http://localhost:8080/api/v1/autonomy/config \
  -H "Content-Type: application/json" \
  -d '{
    "default_level": "supervised",
    "per_action_overrides": {
      "isolate_host": "assisted",
      "create_ticket": "autonomous"
    },
    "per_severity_overrides": {
      "critical": "assisted",
      "low": "autonomous"
    },
    "time_based_rules": [
      {
        "name": "Business hours - supervised",
        "start_hour": 9,
        "end_hour": 17,
        "days_of_week": [1, 2, 3, 4, 5],
        "level": "supervised"
      },
      {
        "name": "Off-hours - autonomous",
        "start_hour": 17,
        "end_hour": 9,
        "days_of_week": [0, 1, 2, 3, 4, 5, 6],
        "level": "autonomous"
      }
    ],
    "emergency_contacts": ["[email protected]"]
  }'

Resolution Priority

When resolving the autonomy level for a given action, overrides are checked in this order:

Per-action overrides (highest priority)
Per-severity overrides
Time-based rules
Default level (fallback)

Resolve for a Specific Action

Check what the system would decide for a specific action + severity combination:

curl -X POST http://localhost:8080/api/v1/autonomy/resolve \
  -H "Content-Type: application/json" \
  -d '{"action": "isolate_host", "severity": "critical"}'

Response:

{
  "level": "assisted",
  "auto_execute": false,
  "reason": "Per-action override for 'isolate_host'"
}

Time-Based Rules

Time-based rules let you run with less autonomy during business hours (when analysts are available) and more autonomy during nights and weekends.

Field	Description
`name`	Human-readable rule name
`start_hour`	Start hour, 0-23 inclusive
`end_hour`	End hour, 0-24 exclusive
`days_of_week`	Array of days (0=Sunday through 6=Saturday)
`level`	Autonomy level when rule applies

Hours wrap around midnight: start_hour: 22, end_hour: 6 means 10 PM to 6 AM.

Execution Guardrails

The guardrails system (configured in config/guardrails.yaml) provides hard limits regardless of autonomy level:

Forbidden actions -- actions that can never be automated (e.g., delete_user, wipe_host)
Protected assets -- targets that always require human approval (production systems, domain controllers)
Rate limits -- maximum actions per hour/day to prevent runaway automation
Blast radius limits -- caps on how many targets a single action can affect

See Guardrails Reference for full configuration details.

Audit Log

Every autonomy decision is logged for compliance and debugging:

curl "http://localhost:8080/api/v1/autonomy/audit?limit=20"

# Filter by incident
curl "http://localhost:8080/api/v1/autonomy/audit?incident_id={id}"

Each audit entry records:

Action and severity evaluated
Resolved autonomy level
Whether auto-execution was allowed
Reason for the decision
Whether the action was actually executed
Execution outcome

Triage Warden