416 tests and a 'destroy all' button: where agent projects break

The author has been running an autonomous agent called Groundhog (Surok) since February 2026 on a research loop with Claude Code and --dangerously-skip-permissions. They then audited six of their own agent projects against a composite checklist from six sources (including Sber's AI-DISRUPT PDLC methodology) and found three recurring failure patterns: safety enforced only in prompts (model can be overridden), irreversible actions without confirmation (e.g., a rebuild command wipes all enriched data), and zero regression tests for behavioral bugs. The author also reviewed an anonymous large open-source agent project: it had code-level permission checks and regression tests for past bugs, but still sent emails irreversibly without a draft/confirmation step. A newer open-source framework (HarnessX) offers an interrupt_on gate but it's optional and not default. Sber's PDLC methodology formalizes many of the missing controls: policy as code, evals as completion contracts, and an R0–R5 permission ladder with mandatory abort/rollback for state-changing operations at R3+. The author built a seventh project implementing some lessons learned: now evals run before each release and incidents become named regression tests.