Skip to content
VB-OSVB-OSVB-OS Docs

Govern AI Agent Actions with VB-OS

Autonomous agents, whether AI-driven, rule-based, or hybrid, take actions with real-world consequences. VB-OS provides deterministic authorization boundaries around those actions.

The Problem

Section titled “The Problem”

Probabilistic guardrails evaluate or filter agent outputs using model-dependent or probabilistic mechanisms. They can reduce undesirable outcomes, but their decisions do not provide deterministic replay guarantees.

VB-OS takes a different approach: separate the agent’s capability to act from the authority to act. The agent produces a recommendation. VB-OS independently determines whether that recommendation is authorized to execute — and that determination is deterministic, replayable, and model-independent.

Integration Pattern

Section titled “Integration Pattern”
┌──────────────┐     ┌──────────────┐     ┌──────────┐
│  Agent       │────▶│  VB-OS       │────▶│  Action  │
│  (recommends │     │  (authorizes │     │  (executes│
│   action)    │     │   or defers) │     │   if     │
└──────────────┘     └──────────────┘     │  ASSERT) │
                                          └──────────┘
  1. The agent produces a recommendation (e.g., “execute trade”, “send notification”, “deploy update”)
  2. Your application submits the recommendation context as evidence to VB-OS
  3. VB-OS evaluates the evidence against the authorization boundary
  4. If ASSERT: the action proceeds
  5. If DEFER: the action is blocked, logged, and optionally escalated

Example Boundary

Section titled “Example Boundary”
boundary_id: agent-action-gate
version: 1
scope: production


require_evidence: agent_id
require_evidence: action_type
require_evidence: risk_level
require_evidence: authorization_tier
require_type: risk_level: integer
require_type: authorization_tier: integer


prohibit_evidence: raw_model_output


predicate: authorized_tier: authorization_tier >= 2
predicate: risk_within_bounds: risk_level <= 50

Implementation

Section titled “Implementation”
from vbos import VBOSClient


client = VBOSClient(api_key="YOUR_API_KEY")


def execute_agent_action(agent_id, action, risk_level, auth_tier):
    result = client.verify(
        workload={
            "agent_id": agent_id,
            "action_type": action,
            "risk_level": risk_level,
            "authorization_tier": auth_tier,
        },
        project="agent-platform",
        boundary_ref="agent-action-gate",
    )


    if result.decision == "ASSERT":
        perform_action(action)
        return {"status": "executed", "evaluation_id": result.evaluation_id}
    else:
        escalate_to_human(action, result.failure_reasons)
        return {"status": "deferred", "reasons": result.failure_reasons}

Key Properties for Agent Governance

Section titled “Key Properties for Agent Governance”
  • Deterministic: the same agent context always produces the same authorization decision
  • Replayable: every authorization decision can be reproduced for audit
  • Model-independent: swap the agent model; the authorization boundary is unchanged
  • Binary: ASSERT or DEFER, no confidence scores or partial approvals
  • Fail-closed: missing context produces DEFER, not a best guess

Next Steps

Section titled “Next Steps”