Step 1
AI generates action
Runtime authorization for AI agents
An AI agent proposes an action, BobSentry evaluates it before execution and returns ALLOW / REQUIRE_APPROVAL / BLOCK, and every decision is captured as signed decision evidence.
BobSentry sits at your agent's action execution boundary — in front of APIs, databases, email, and workflow side effects. In the default metadata-only integration, raw payloads stay inside your environment; only normalized classification metadata reaches BobSentry.
How it works
Step 2
BobSentry runtime evaluates and signs
Step 3
Policy enforced before execution
AI -> BobSentry Runtime -> Your Systems
Core capabilities
Action Validation Gate
- Evaluate every action before execution
- Validate actor, target, and workflow intent
- Return ALLOW, REQUIRE_APPROVAL, or BLOCK
Approval Escalation Controls
- Escalate high-impact actions to operators
- Capture approver, justification, and status
- Prevent unsafe auto-execution paths
Runtime Evidence Continuity
- Attach evidence IDs to each decision event
- Preserve timestamps and workflow source context
- Exportable JSON/PDF evidence bundles
Observe to Enforce Rollout
- Start in Observe to baseline workflow behavior
- Move to Enforce when controls are production-ready
- Apply mode per agent and workflow risk profile
Execution Decision Engine
Each action is evaluated in runtime context before execution.
Allowed ALLOW
BobSentry policy permits this action.
Example trigger
- Low-risk action within policy
Requires Approval REQUIRE_APPROVAL
BobSentry requires authorization before the integration should proceed.
Example triggers
- Destructive action
- Elevated-risk context
Blocked BLOCK
BobSentry policy denies this action.
Example trigger
- Policy violation
Critical actions should never auto-execute
High trust alone is not enough for destructive execution. BobSentry can escalate, require approval, and enforce runtime policy controls for high-impact operations.
DELETE_DATABASE
TRANSFER_FUNDS
External PII exposure
Bulk workflow execution
Runtime authorization flow
AI Agent
BobSentry Runtime Evaluate API
ALLOW / REQUIRE_APPROVAL / BLOCK
API / DB / Email / Workflow
Works with your existing agents
Integrates at your agent's tool or action boundary through the BobSentry API or SDK.
LangChainOpenAI AgentsCrewAIMCP workflowsCustom agents
Preview the runtime decision engine
Run deterministic no-signup scenarios before creating a sandbox key.
No signup required. Public demo uses deterministic decision artifacts for stable evaluation.
Browser demo: POST /api/v1/ai/proxy/demo
Demo endpoint — no signup required, runs entirely on metadata.
For production: generate a sandbox key and POST metadata to https://runtime.bobsentry.com/v1/runtime/evaluate. In the default metadata-only integration, raw payloads stay inside your environment and raw PHI is not sent to BobSentry.
1. Choose a scenario →2. Preview Runtime Engine
Action logs
Seeded operational events show runtime decisions, escalation boundaries, and evidence references.
| Time | Workflow Source | Action | Decision | Risk Level | Reasoning | Evidence ID |
|---|---|---|---|---|---|---|
| 10:37:12 AM | healthcare_reporting_agent | send_external_email | BlockedBLOCK | HIGH | Outbound email contains patient identifiers and target domain is not approved. | evd_health_email_9f31 |
| 10:34:40 AM | crm_ops_automation | bulk_crm_export | Requires ApprovalREQUIRE_APPROVAL | HIGH | Bulk export exceeds customer data threshold and requires supervisor approval. | evd_crm_export_63ab |
| 10:31:08 AM | release_orchestrator | kubernetes_prod_deploy | Requires ApprovalREQUIRE_APPROVAL | MEDIUM | Deployment targets production namespace outside approved maintenance window. | evd_deploy_gate_42cd |
| 10:28:22 AM | support_automation | support_refund_payout | BlockedBLOCK | HIGH | Payout amount and account change pattern match fraud controls. | evd_finance_payout_a1d0 |
| 10:25:15 AM | ops_status_assistant | internal_status_email | AllowedALLOW | LOW | Internal communication matches allowed channel and low-risk content profile. | evd_internal_email_73ef |
Use cases
AI customer support agents
AI copilots
Internal automation tools
Fintech / healthcare AI
Operational trust signals
- Action decisions include risk level, reasoning, timestamp, and evidence reference.
- Metadata-only transport keeps raw payload content inside your environment.
- Execution controls sit in front of APIs, email, database, and workflow side effects.
Enforcement Mode
Observe
See what AI would do (no blocking)
Enforce
Actively block unsafe actions
Observe logs actions without blocking execution.
Operational outcome
- Access is not execution: every high-impact action receives a runtime decision first.
- Escalation and approvals preserve bounded autonomy for autonomous workflows.
- Signed runtime evidence preserves accountability for operators and auditors.