This is a Meta Superintelligence Labs paper proposing an inference-time harness that splits an agent into workers and a controller. The controller runs a four-stage cycle — assess state, propose next computations, evaluate them against remaining budget, dispatch workers with selected artifact context — and keeps only a compact self-authored state between cycles instead of replaying full history. Everything lives in persistent memory as an artifact graph. The claim: control decisions are a distinct capability from object-level work, and making them explicit beats interleaving them into the transcript. Numbers: on ProgramBench with GPT-5.5 at 1200 model calls, 71.5% versus 58.0% for Codex and 63.7% for a matched Direct Control Agent with the same workers and budget. Mean gains of 3.6–4.2 points across four benchmarks, three frontier models, all 12 matched comparisons. Controller calls count against the same budget, so the deliberation has to pay for itself. The honest caveats are in the paper itself: it loses at small budgets (Opus 4.8, 400 calls: 56.6% vs 62.7%), the Direct Control comparison conflates staging, state compression, and memory access, and worker self-confidence ranks near chance (Type-2 AUC 0.55) while the controller's verdict hits 0.88. Coverage gains are real; final selection gains are patchy. Practical takeaway: the interesting move is charging the controller against the same budget and still winning. Most "orchestrator" papers quietly exempt their own overhead.