USE CASE 05 / AGENTIC-SIDECAR
Before the call.
Keep intent in control.
Watch an agent work across external services, MCP, a vector database and an action tool. Sidecar supervises every proposal before the call is released.
SUPERVISED EXECUTION / AGENTIC-SIDECAR
One agent. A checkpoint before every call.
Resolve order NS-1042: retrieve only the relevant customer and policy context, preserve personal data, and propose an eligible refund only when policy permits.
Context attachedSupport agent
Propose a scoped customer lookup.
PREPARING PROPOSALGuardrails
Check the proposed call against action policies.
WAITINGIntent & context
Restore the goal and constrain the call’s scope.
WAITINGSelf-critique
Review the proposal, evidence and expected risk.
WAITINGHuman approval
Request approval only for critical actions.
WAITINGCarry the original goal and the latest result into the next proposed call.
The agent prepares a call.
The proposed customer lookup is held before any external request is made.
- Propose
- Guardrails
- Context
- Critique
- Approval
- Execute
- Result
- Re-inject
Concept simulation: only critical actions in this example require human review. “Human approval” is a scripted event, not a real person responding. Actual approval requirements depend on configured policy. No tool, model, database or infrastructure call is made.
SUPERVISION TRAVELS WITH THE AGENT
Guardrails around the action. Intent throughout the loop.
Each new proposal returns through the same checks. Context is restored before the agent moves on, critique keeps the action aligned with its goal, and the execution gate stays held until automated checks pass and any required human review is complete.
Explore the Control Tower fleet demo →