From swarmbench's migrate-upgrade-oscar-django benchmark at size L. The code is django-oscar 1.6.7 (BSD-3-Clause); the swarm got the whole checkout and a public checker, tools/check.py.

It is graded by Oscar 1.6.7's own tests that the agents never saw (935 of them), behind a static gate: any removed Django API left in the code, a broken manage.py check, or a missing migration scores 0. The hidden test names stay private so the benchmark stays a test; Raw results has the failures counted by app.

The agents did not choose their strategy alone: the root got the task with our operating instructions (write shared conventions, delegate one lead per group, leads delegate one agent per file, review, fix). Those instructions are in the root's first message and the role briefs it read; the harness's role preamble and internal ids are left out of the transcripts below; everything else is as it happened.

The root's last full check did not run the tests. A bug in the benchmark's own checker (tools/check.py, fixed since) ran every check after the first without Django, so check.json in Raw results reports the static checks only, with a note the root added. The root therefore never saw the failing tests, and this run did not exercise its fix-the-failures step.

795 of 935 hidden tests pass (85%) and the static gate passed; the benchmark scores this PARTIAL. It is the second of two comparable runs (the first passed 787, 84%). There are no single-agent or smaller-swarm runs of this task on the same setup yet, so it does not show that a swarm migrates better than one agent. 49 minutes, $8.10 in total ($1.28 model, $6.82 sandboxes), model gpt-5.6-luna.

Models: gpt-5.6-luna.