I Let AI Touch Production, With a Kill Switch

Traditional software follows code paths engineers already defined. An agent can choose its next action at runtime. That flexibility is useful, but it creates a gap: ordinary application permissions are often too broad, while a simple confirmation dialog gives too little context.

I needed the system to answer five questions clearly: who is asking, which tool do they want, what data will change, does a person need to approve it, and can the action still be stopped?

I designed the prototype around one rule: the model never talks directly to an operational API.

The agent submits a structured action proposal to a Laravel control layer. That layer checks the actor, environment, tool, requested fields, and action risk. Access is default-deny. Low-risk reads may continue; sensitive writes enter an Angular approval queue that shows exactly what the agent wants to do and what payload will be sent.

The design adds schema validation, field allowlists, timeouts, rate limits, and an audit trail connecting the proposal, policy decision, reviewer response, and tool result. Global and scoped kill switches can stop all new execution or disable one agent, tool, or workflow without taking the rest of the application down.