We present a governance framework for autonomous AI agents that build and operate production software. The framework emerged from operating ten AI agent personas to construct a 62-table SaaS platform with billing, authentication, and multi-tenant isolation. We identified eight failure modes absent from traditional software development: context rot from conversation history compression, unbounded fix-break retry loops, cross-tenant query omissions, mutable financial records, non-idempotent writes, systematic self-assessment bias in code review classification, builder-auditor conflation when the same agent builds and reviews its own work, and model-capability mismatch when implementation-optimized models perform judgment tasks. We address each through structural mechanisms that the agent cannot circumvent: workflow phases enforce fresh-context boundaries so degradation cannot propagate across roles, a classification script assigns review tiers based on diff properties rather than agent self-report, errors are classified into four types before recovery begins with hard per-category iteration limits, multi-agent decisions use sequential written artifacts rather than group deliberation, financial tables enforce append-only semantics with reversal-entry corrections, all write operations require duplicate-call idempotency tests, model-specific capability requirements are enforced per workflow phase, and an asymmetric audit methodology requires grading before fixing with a published scorecard. The mechanisms are organized into seven process layers and a five-phase deterministic workflow, then packaged as a transferable template. We describe the design, its rationale, and its relationship to existing work in agent safety and software process frameworks.
Greg Arnold (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: