Raven is an open-source multi-agent orchestration framework (EverMind AI, arXiv 2609.33439) whose premise is that an agent's capability lives in the model–harness pair, not the model alone. A Host Agent decomposes a goal, matches subtasks to registered agents — native specialists (Research, Code, Design, Oncall) plus third-party ones like Claude Code and Codex — and executes them as a typed DAG with an append-only artifact ledger for handoffs. Harnesses evolve via a "gene bank" and gated selection; memory and Skill Forge persist and reuse experience across tasks. The theory section gives sufficient conditions (contract realization, compatible plans, bounded error) under which composition beats any single agent under the same budget — honest, if the premises hold, which averages on a benchmark don't establish. Headline numbers: 10.4–10.5 pp Exact Match gains on their own MAOB benchmark, and Raven-Oncall at 82.35% success for $5.10/case versus baseline 64.71% at $13.60. The part I keep re-reading is the claim that harnesses autonomously construct and evolve other harnesses. That's a harness whose job is harnesses, which by induction should eventually be run by a harness. Somewhere down that stack is a 300-baud acoustic coupler doing the actual work, as it always has. The evals are self-conducted, the benchmark (MAOB) is theirs, and the theory explicitly warns that measured success rates don't prove the reliability premises. Treat it as an architecture worth studying, not a capability claim worth betting on yet. Open question: do the DAG-exactness gains survive on third-party graphs, or only on plans that look like the authors' own decomposition habits?