WorkflowOps is a multi‑agent workflow orchestration framework that learns collaboration priors from historical workflows and expands its agent pool on demand to satisfy new capability requirements. The system consists of three coupled mechanisms.\ \ The first mechanism builds a transition probability matrix $P_{ij}$ that records how often agent $i$ hands off to agent $j$ in past executions. During DAG construction this matrix provides soft guidance: intra‑layer ordering optimization improves overall success, probability‑thresholded edge suggestion keeps only high‑confidence transitions, and transitive reduction maximizes parallelism.\ \ The second mechanism is a sufficiency‑driven agent creation loop. Semantic matching scores expose capability gaps; an LLM then generates specialized agents, which are immediately injected into the collaboration matrix so that the new agents inherit predicted priors and become usable without delay.\ \ The third mechanism adopts a layered semantic matching strategy. A fast deterministic pass uses pre‑trained sentence embeddings for capability matching; only low‑confidence cases trigger LLM verification, cutting pure‑LLM routing calls by over 80%.\ \ Experiments on mixed code, math, and QA suites show that WorkflowOps improves end‑to‑end pass rates compared to recent workflow‑construction baselines, with the largest gains on structured, decomposable tasks where past handoff patterns transfer effectively.\ \ Review