Better Agent Tooling Can’t Hide Near‑Zero Success on Real Tasks

Cua’s Quad Driver moves computer‑use agents from screen‑takeover to background operation across macOS, Windows, and Linux. In their benchmarks, switching from a built‑in tool raised agent success from 62% to 80% and cut token usage by 34%, largely by focusing on a single window instead of the full desktop.

Companion Kua Bench shows that on electrical‑engineering tasks with real professional software, the best agent passed only 6/25 tasks, and success dropped to 0% from a blank schematic. The leaderboard is flat, with no model exceeding 30% reward, and tooling gains don’t compensate for inadequate agent intelligence on creative or unstructured tasks.

For RL training, Kua Fleet uses a demand‑based sandbox pool to hide startup latency and maximize expensive GPU utilization. This shifts infrastructure cost from idle GPUs to cheaper pre‑warmed sandbox instances. The approach is pragmatic but unproven at scale outside the vendor’s own pipelines, and reliance on undocumented OS APIs introduces maintenance risk.

Engineering leaders should treat this as a signal that computer‑use infrastructure is maturing, not that agents are production‑ready. The real decisions revolve around investing in driver‑level integration, robust evaluation frameworks, and efficient training orchestration, while accepting that complex task automation remains a research problem.