{
  "name": "SatGate AI Agent Runaway Spend Index",
  "month": "2026-04",
  "published": "2026-04-26",
  "methodology": "Modeled runaway spend exposure for autonomous AI agents, MCP tools, retry loops, and delegated sub-agent fanout. Values are scenario estimates for category benchmarking, not customer telemetry.",
  "summary": {
    "medianUncontrolledIncident": 1840,
    "p90UncontrolledIncident": 18480,
    "largestModeledIncident": 134400,
    "medianAvoidedWithRequestPathControls": 94.2,
    "topFailureMode": "retry storms against paid MCP tools and model APIs"
  },
  "scenarios": [
    { "failureMode": "OpenAI retry loop", "uncontrolledCostUsd": 7200, "controlledCostUsd": 250, "avoidedPercent": 96.5 },
    { "failureMode": "MCP browser automation loop", "uncontrolledCostUsd": 1840, "controlledCostUsd": 120, "avoidedPercent": 93.5 },
    { "failureMode": "sub-agent research fanout", "uncontrolledCostUsd": 18480, "controlledCostUsd": 900, "avoidedPercent": 95.1 },
    { "failureMode": "paid data API polling loop", "uncontrolledCostUsd": 9600, "controlledCostUsd": 600, "avoidedPercent": 93.8 },
    { "failureMode": "multi-tenant agent swarm", "uncontrolledCostUsd": 134400, "controlledCostUsd": 6000, "avoidedPercent": 95.5 }
  ],
  "recommendedControls": [
    "per-agent daily and session budgets",
    "per-request maximum cost",
    "MCP tool price registry",
    "delegation and sub-agent fanout caps",
    "revocable capability tokens",
    "budget exhaustion errors agents can reason about",
    "request-path enforcement before upstream calls execute"
  ]
}