Enterprise Agent Governance 🟡 BETA
Centralized monitoring, token auditing, and safety guards for AI agents
Overview
Enterprise Agent Governance provides administrators with a centralized control room to observe, audit, and secure all AI agents operating across the workspace. It combines real-time telemetry, dynamic security guards, script execution isolation, and a live dashboard with kill-switch capabilities.
Architecture
┌─────────────────────────────────────────────────┐
│ Governance Dashboard (HTMX) │
│ ┌──────────┐ ┌──────────┐ ┌──────────────────┐ │
│ │ Live │ │ Token │ │ Kill Switches │ │
│ │ Metrics │ │ Auditing │ │ (bot-level) │ │
│ └──────────┘ └──────────┘ └──────────────────┘ │
└──────────────────────┬──────────────────────────┘
│ SSE / WebSocket
┌──────────────────────▼──────────────────────────┐
│ Metrics Ingestion Pipeline │
│ ┌──────────┐ ┌──────────┐ ┌──────────────────┐ │
│ │ LLM Call │ │ Tool │ │ Script │ │
│ │ Intercept│ │ Execution│ │ Guard │ │
│ └──────────┘ └──────────┘ └──────────────────┘ │
└──────────────────────────────────────────────────┘
Telemetry Ingestion
A middleware intercepts all LLM calls and tool executions across the system, ingesting:
| Metric | Description |
|---|---|
| Timestamp | Event time |
| Bot ID | Which bot triggered the event |
| Session ID | User session context |
| Model | LLM model used |
| Prompt tokens | Input token count |
| Completion tokens | Output token count |
| Tool name | Executed tool/script |
| Latency | Execution duration |
Metrics are persisted to the database and pushed to Valkey for live dashboard rendering via SSE.
Dynamic Security Guards
The script guard (script_guard.rs) inspects Rhai BASIC scripts before execution, detecting 17+ dangerous patterns:
| Pattern | Detection |
|---|---|
| Secret key leakage | Reads from /tmp/vault-token-gb, env vars |
| Shell injection | Command::new(), shell execution |
| File system escape | Writes outside .gbdrive/ |
| Network probing | Internal IP/port scanning |
| Resource exhaustion | Infinite loops, excessive allocations |
Blocked events are logged to an immutable security_audit_logs table.
Administrator Console
Dashboard
The governance dashboard at /suite/governance provides:
- Live Metrics: Real-time WebSocket throughput and token consumption costs grouped by bot and department
- Kill Switches: Immediately disable specific bots or revoke external tool access
- Security Incidents: Real-time alarms via Server-Sent Events (SSE)
- Audit Log: Searchable history of all security events
Kill Switch
When a bot is disabled via the kill switch:
- WebSocket connections for that bot are rejected
- Pending tool executions are cancelled
- An incident is logged to the audit table
- Administrators are notified via the dashboard
Configuration
| config.csv key | Description | Default |
|---|---|---|
governance-enabled | Enable governance features | true |
governance-audit-log | Enable immutable audit log | true |
governance-alert-email | Email for critical alerts | (none) |
Feature Flag
Enable with monitoring feature flag:
botserver = { features = ["monitoring"] }
Security
- Metric ingestion failures never block bot conversation paths — background threads with bounded channels
- User message content is redacted from monitoring console logs unless detailed auditing is explicitly enabled
- Kill switch fail-safe: DB pool overflow falls back to local file logging