Agentic Engineering Framework

About

AI coding agents fail in predictable ways: they edit without a task, destroy files with force flags, and run out of context mid-change. The framework interposes itself between the agent and the repository using a PreToolUse hook, tiered authority rules (human sovereignty, framework authority, agent initiative), and Markdown task files with YAML frontmatter carrying acceptance criteria and verification commands. A budget gate watches the live transcript and blocks source edits when the context window nears its limit, forcing a commit and handover instead of a truncated change. Memory, component maps, and audit checks (over 260 checks run on every push and on a 30-minute cron) provide continuity and traceability across sessions. It is used by developers running Claude Code who want audit-grade accountability; the framework itself is developed under its own governance, with thousands of self-governed commits.