AI Agent & Reasoning Plane Spec
Deterministic protocol specification, JSON schemas, Model Context Protocol (MCP) tool catalogs, and Risk Tier classification taxonomy for autonomous agents.
Rule of Decoupled Intelligence for Autonomous Agents
AI agents MUST NOT generate raw SQL, execute bash commands, or directly invoke infrastructure endpoints. Agents ONLY emit schema-validated ActionIntent payloads. The Action Gateway independently evaluates identity, tenant boundary, blast radius, and risk tier before any execution occurs.
Interactive Action Gateway Simulator
Experience the strict Decoupled Intelligence boundary: AI proposes an ActionIntent, the Action Gateway enforces deterministic risk tiers, cryptographic HMAC signing, and fails closed.
Autonomous agent assesses system state and formulates a strict, declarative ActionIntent.
Evaluates blast radius, computes HMAC tamper seal, and maps action to policy safeguards.
Executes only with valid cryptographic token; records tamper-proof ledger entry.
Risk Classification Matrix
The Action Gateway deterministically maps every ActionIntent into one of four Risk Tiers based on verb semantics, environment, and blast radius.
| Tier | Classification | Allowed Verbs | Gatekeeper Requirement | Agent Behavior |
|---|---|---|---|---|
| Tier 0 | Read-only / Observability | get, list, read, query, logs, metrics, status | Immediate execution + signed audit log | Synchronous result returned immediately |
| Tier 1 | Low-Mutation / Safe Addition | create, add_dns_record, set, upload, stage, preview | Diff preview + confirmationToken verification | Inspect diff, present to user, re-submit token |
| Tier 2 | High-Mutation / Service Deploy | deploy_worker, scale, restart, switch_traffic, update_dns_zone | Step-up 2FA / TOTP authentication challenge | Prompt human operator for 2FA verification token |
| Tier 3 | Destructive / Irreversible | drop_table, delete_database, purge_all_dns, delete_bucket | Dual-Authorization Human Staff Approval Queue | Held in queue. Agent monitors status via requestId |
Canonical ActionIntent JSON Schema
{
"$schema": "http://json-schema.org/draft-07/schema#",
"title": "ActionIntent",
"type": "object",
"required": [
"intentId",
"actor",
"targetService",
"actionVerb",
"resourceUri",
"parameters",
"idempotencyKey",
"createdAt"
],
"properties": {
"intentId": { "type": "string", "format": "uuid" },
"actor": {
"type": "object",
"required": ["id", "type", "tenantId", "roles"],
"properties": {
"id": { "type": "string" },
"type": { "type": "string", "enum": ["ai_agent", "user", "system"] },
"tenantId": { "type": "string" },
"email": { "type": "string", "format": "email" },
"roles": { "type": "array", "items": { "type": "string" } },
"delegationToken": { "type": "string" },
"clientIp": { "type": "string" }
}
},
"targetService": {
"type": "string",
"enum": ["compute", "storage", "database", "dns", "ai_gateway", "network", "billing", "iam"]
},
"actionVerb": { "type": "string" },
"resourceUri": { "type": "string" },
"parameters": { "type": "object" },
"idempotencyKey": { "type": "string", "minLength": 8 },
"reasoningTrace": { "type": "string" },
"confidenceScore": { "type": "number", "minimum": 0, "maximum": 1 },
"createdAt": { "type": "string", "format": "date-time" }
}
}Model Context Protocol (MCP) JSON-RPC 2.0
All tools exposed by VaultHost conform to the standard MCP specification. External LLM orchestrators can invoke tools directly against /api/action-gateway/mcp.
{
"jsonrpc": "2.0",
"id": "req-001",
"method": "tools/call",
"params": {
"name": "vaulthost_workers_deploy",
"arguments": {
"workerName": "edge-auth-router",
"action": "deploy",
"environment": "production",
"scriptCode": "export default { async fetch(req) { return new Response('OK'); } }"
}
}
}