“Agents do real work in our systems now. Knowing every tool call is checked against a policy is what let us take them out of the sandbox.”
Daniel Okafor
VP Engineering, Arcflow
01 // OVERVIEW
Autonomous agents outgrew their sandbox
Arcflow built agents that booked freight, adjusted routing and updated records without a person in the loop. In testing that was useful; in production it meant software could move money and change customer commitments.
Tool access had been granted broadly during prototyping and never narrowed.
THE CHALLENGE
Agents could reach tools and take actions well beyond the scope anyone had approved.
02 // Objective
Give every agent a boundary it cannot cross
Arcflow wanted the agents in production, not paused, so the objective was a control layer rather than a review board.
Every agent needed an explicit scope, every high-impact action needed a check, and anything unexpected needed to be reversible.
03 // Solution
Permission controls, action review and rollback
1. Permission Controls
Each agent received a least-privilege scope tied to its purpose, so a routing agent could not reach billing and a support agent could not write to inventory.
2. Action Review
High-impact calls were evaluated against policy before execution, with the riskiest routed to a human for a one-click approval.
3. Rollback
Every action was recorded with enough context to be reversed, which is what gave the team the confidence to widen the remit of the agents.
