A services-firm architecture for long-horizon agentic builds. In production across six engagements, two of them our own systems, reported with an honest defect analysis and a cost-of-ownership benchmark.
Long-horizon software builds, spanning days of autonomous execution and months of equivalent human effort, remain an open challenge for agentic AI. We argue the binding constraint is not model capability but operational architecture, and that operating an agent harness inside a services firm with contractual delivery accountability yields a different and more tractable design space than shipping the same harness as a developer product.
We describe Blueprint, a context operating system producing a gated chain of specification artifacts; Agentic Builder, a model-agnostic multi-agent execution harness; and Compass, a structured knowledge layer that accumulates competence across engagements. Excavator, a second Blueprint ingestion mode, recovers requirements from existing legacy systems so the same machinery rebuilds them.
Our central contribution is the identification of five design constraints that change when an agent harness is operated by a services firm rather than shipped as a product, together with the architecture that satisfies them. We report empirical results, including a total-cost-of-ownership benchmark and a full defect analysis that surfaces specification-detail thresholds and test-automation integration as the two dominant open problems.
The paper is in preparation, with publication expected in 2026. Engagement-level figures are being finalized before submission.
We can walk your team through the architecture before the paper ships.