Autonomous agents need brakes: tool use with human-in-the-loop
Agent workflows create leverage only when refunds, legal language, and VIP accounts stay behind explicit gates. Here is the control plane we ship with every production agent.
Julian Park · Principal AI Engineer · 8 min · June 18, 2026
An agent that can call tools is an employee with a very fast keyboard. You would not give a new hire refund authority on day one. Do not give it to a model because the demo looked smooth.
We design agents as graphs: retrieve, decide, act, write back. Each act node has a policy. Low-risk actions (order status, password reset links, knowledge answers with citations) can complete. High-risk actions pause for a human with the full trace attached.
Observability is the product. Every tool call, retrieved chunk, and token cost should land in a trace you can replay. When something goes wrong (and it will), the question is not “did the model misbehave?” It is “which policy failed, and how do we prevent the class of failure?”
Cost control belongs in the same dashboard. Agents that loop, over-retrieve, or call the largest model for a FAQ will quietly become the most expensive intern in the building. Route by task difficulty. Keep a small model on the well-worn paths.