Production AI agents
Point of view
Every production agent needs an explicit contract for what it may know, what it may do, what evidence it must leave and how a person can stop it.
Establish identity
Give the agent a distinct, narrow and auditable identity with short-lived credentials. Permissions should reflect its job, environment and current task.
Identity makes revocation, attribution and least privilege possible when actions span systems.
Intercept action
Inspect requests and proposed actions for prompt injection, sensitive data, policy violations and business limits before execution. High-consequence actions should pause for approval.
Output filtering alone is insufficient when an agent can write, deploy, purchase, message or delete.
Operate the agent
Record plans, tool calls, approvals, failures and outcomes in a privacy-conscious trace. Evaluate task completion alongside safety, cost, latency and human intervention.
Provide a kill switch, graceful degradation and a clear human escalation path.
Design principles
What changes Monday
Inventory every tool an agent can invoke.
Replace inherited credentials with task-scoped access.
Define approval thresholds and a kill path.
Run adversarial scenarios before granting production authority.
AI agents become trustworthy when the system makes authority visible, limited and reversible.
Discuss this perspectiveContinue exploring
AI can accelerate delivery only when intent, boundaries, evidence and failure behavior are explicit enough to govern the work.
Read field note
Generated tests create value when they challenge behavior, boundaries and failure using an independent definition of correctness.
Read field note
The safest use of coding agents in mature systems begins with context recovery, a focused change and a tested path back.
Read field note