From One Factory to a Two-Story Harness: How a Dev Pipeline Grew in Under a Year

In November of the previous year, the author began assembling a multi-agent development environment – not a wrapper over a single model but a factory that takes a project from requirements to working code without breaking what must not be broken. Starting with Cursor and rule-based agents, the system used a process approach: each stage had its own role-subagent and reviewer. Within months, the need for multiple parallel pipelines emerged – for marketing, product, design, content. A top-level control plane was added: a root router, control plane (239 Work Items), a harness learning layer (ratchet, sensors, rule-evals). The factory became two-story: cross-cutting control above, local gates inside each factory. The author practices ratchet – every agent failure is turned into a permanent environment fix, not a chat note. Sensor-first means a new invariant starts as a detector script with PASS/FAIL contract before becoming a text rule. Measurability ensures a rule changes behavior – a gate passes good artifacts and blocks violators. The author is not locked into Cursor: open standards (Skills, AGENTS.md, MCP) are supported by Claude, Codex, and Gemini. A minimal working skeleton is provided at agent-harness-reference. Some invariants still lack detectors and rely on review discipline.