Type: url
Original: https://arxiv.org/pdf/2602.20478
Status: generated
Summary
“Codified Context: Infrastructure for AI Agents in a Complex Codebase” is a systems paper and experience report by an independent researcher describing a three-tier context infrastructure built while developing a 108,000-line C# distributed system with Claude Code as the sole code-generation tool: an always-loaded project constitution (hot memory), 19 specialized domain-expert agents, and 34 on-demand specification documents served through an MCP retrieval server (cold memory). It reports growth and interaction metrics across 283 development sessions and four observational case studies of codified context preventing failures and keeping behaviour consistent across sessions, and names specification staleness as the primary failure mode. The author stresses that the evaluation is single-developer, single-project and observational, with no causal claims.