An 'agent' is just an LLM that can call tools in a loop. That is powerful and dangerous in equal measure. The difference between a useful agent and a liability is not the model - it is the guardrails around it.
Give tools narrow, typed interfaces
def refund(order_id: str, amount_cents: int) -> str:
"""Refund up to 5000 cents. Larger refunds require human approval."""
if amount_cents > 5000:
return queue_for_human_approval(order_id, amount_cents)
return payments.refund(order_id, amount_cents)The four choices that make an agent safe
- A named owner responsible for the business outcome
- Least-privilege tools - it can only do what it is scoped to do
- Human approval for anything irreversible (money, prod, customer data)
- A log of every action for audit and debugging
# log every tool call for an audit trail
log.info('agent_action', tool=tool_name, args=args, user=ctx.user, approved=approved)An agent is a workflow with accountability. Remove the accountability and you have not automated the work - you have hidden it.
Start with one workflow, give the agent the smallest possible set of guarded tools, and keep a human in the loop for consequential actions. Expand scope only as trust is earned.