Red-team a fine-tuning plan with parallel skeptic agents

D'Erme was planning FT-4, a LoRA and GRPO fine-tune of Qwen3.5-9B meant to use his own memory system, tested at 4 bits on an old 8 GB GTX 1070 Ti. Before paying for training GPUs he asked one Claude Code session, running Opus 5.5 at xhigh effort with ultracode on, to play devil's advocate. Ultracode turns a prompt into a workflow script that fans out subagents: here read-only readers over his records, harness code, memos and papers, ten skeptics told to refute each group of claims, and a critic that checked the review against his brief. He then asked for separate sessions per issue, and the coordinator started eight sessions, wrote their briefs, agreed which session owned which shared file, relayed his answers and handed out rented GPUs from vast.ai. All API spend went through one OpenRouter key he capped at USD 50.

The first verdict came back in two hours: GO WITH CHANGES, with ten blockers. Skeptics confirmed no GRPO reward code existed, so training would have scored every answer zero, and readers found his untrained harness already inside the range predicted for the trained model, which a later GPU run confirmed. Over about 45 hours it ran four workflows and nine sessions, one of which started a tenth, while he typed roughly 20 turns. The sessions also built a scorer and reward module with 189 new tests and a fix for 4-bit serving. A later planning workflow corrected 26 claims and added 18 missing items.

He did not measure token spend, so there is no cost figure, and every subagent uses its own context, so a workflow costs far more than one careful turn. The coordinator ran out of context on day two and continued from a summary. Spending approvals did not pass between sessions: one refused to spend on a yes relayed by a peer. He was the only approver for all ten sessions.