Permissions answer
What can this agent access?
Runtime authorization for AI agents
Decide what AI agents should be allowed to do — before they act.
Evaluate consequential actions at runtime. Allow, require human authorization, or deny by policy — with signed decision evidence. Permissions alone are not enough.
BobSentry returns the authorization decision; the integration must honor it. BobSentry does not execute the downstream action.
LIVE API EXAMPLERuntime evaluate
POST https://runtime.bobsentry.com/v1/runtime/evaluate
{
"action": "deploy_production",
"metadata": {
"environment": "production",
"target": "github-deploy-workflow"
}
}
→ { "decision": "REQUIRE_APPROVAL" }Missing layer
Permission is not authorization
Tool or API permission tells you what an agent can access. Runtime authorization determines whether this specific consequential action should proceed now.
BobSentry answers
Should this action proceed right now?
- An agent may legitimately have access to a capability and still need a runtime decision for each use.
- Authorization should evaluate the specific action, target, context, and policy—not a standing grant alone.
How it works
How BobSentry works
BobSentry sits at the decision boundary between agent intent and consequential execution. It evaluates policy, requires human authorization when needed, and records signed evidence. The integration performs any downstream operation.
Agent requests consequential action
Runtime
BobSentry evaluates policy
Decision
ALLOW
ALLOW
Policy permits the requested action
REQUIRE_APPROVAL
REQUIRE_APPROVAL
Human authorization required before proceed
BLOCK
BLOCK
Policy denies the requested action
Authorization
Human authorization where required
Execution
Integrator executes when authorized
Evidence
Signed decision evidence
Authorization before execution. BobSentry does not execute the downstream action—the integrator does, only when policy allows.
Demo
See runtime authorization in action
Watch AI agent authorization in practice: an agent requests a consequential action, BobSentry evaluates policy in real time, requires human approval for AI agents when needed, and creates signed decision evidence before execution can proceed.
BobSentry provides the authorization decision. The integration remains responsible for execution.
The same authorization pattern shown here protects BobSentry's own production deployment pipeline.
Real production workflow
BobSentry protects BobSentry
Our GitHub Actions → AWS production deployment workflow runs through BobSentry before infrastructure mutation. When deploy_production is requested, policy returns REQUIRE_APPROVAL. A human authorizes the action, GitHub Actions performs the deployment, and the execution outcome is recorded separately from the authorization decision.
- deploy_production
- REQUIRE_APPROVAL
- Human authorization
- GitHub Actions deploys to AWS
- Execution outcome recorded
BobSentry authorizes the action. GitHub Actions and AWS perform the deployment.
Runtime decisions
Three outcomes for every consequential action
ALLOW
Policy permits the requested action.
REQUIRE_APPROVAL
Human authorization is required before the integration should proceed.
BLOCK
Policy denies the requested action.
Consequences
Consequential actions need a runtime decision
These are the kinds of operations teams often refuse to let agents run fully autonomously.
- deploy_production
- submit_invoice
- grant_admin_access
- export_patient_records
Product
What you operate in BobSentry
Register agents, define policy, authorize when required, and keep signed evidence of each decision. Use the Simulator to test policy without executing an action.
Agents
Register agent identities and issue Agent-bound runtime credentials that are evaluate-only and revocable.
Policies
Define what may proceed automatically, what requires authorization, and what should be denied—including safe unmatched-action fallbacks.
Approvals
Route consequential actions to a human when policy returns REQUIRE_APPROVAL. Approval authorizes the action; it does not execute it.
Evidence
Keep signed authorization evidence: agent, action, policy, whether authorization was required, who authorized, and when. Execution outcomes are recorded separately from authorization.
Simulator
Test how policy would decide for a proposed action without operational side effects.
Developer path: Developers. Full product overview: Product.
Trust boundary
Authorization boundary, not silent execution
Metadata-only by default: raw business payloads can remain in your environment while BobSentry receives the normalized action context needed for authorization. BobSentry returns a signed authorization result. Starter controls and unmatched-action fallbacks keep unknown actions from falling through to a silent ALLOW.
Customer environment
- AI agent
- Sensitive payload
- Local classification / redaction
Normalized action metadata
BobSentry
- Deterministic policy evaluation
- Human authorization
- Signed decision evidence
COOPERATIVE
The integration calls BobSentry before the consequential operation and is responsible for honoring the authorization decision.
OBSERVE
BobSentry evaluates and records what policy would decide, but does not control downstream execution.
Customer-facing modes today are COOPERATIVE and OBSERVE. There is no customer-selectable ENFORCED mode. Read the full security model on Security.
Working with teams operating consequential agents
We collaborate closely with teams already building or running AI agents that can change production systems, move money, send customer communications, or touch sensitive data.
What actions does your team still refuse to let agents execute automatically?
- Production changes
- Financial operations
- Customer communications
- Access and permission changes
- Sensitive exports
- Destructive operations
- AI agent platforms
- AI automation companies
- Enterprise AI teams
- Healthcare AI startups
- Finance AI systems
Prefer email?
Share the consequential actions you need to authorize and your integration call site.
By contacting us you agree we may use your message to respond. See our Privacy page.