LibreWeddingPlanner maintainer reports skill erosion after multi-agent AI workflows replaced TDD and code review

The author banned AI contributions to LibreWeddingPlanner yet enabled VS Code enhanced autocomplete and multi-worktree agents at work. Tasks once requiring days completed in hours. TDD invariants collapsed when agents wrote both tests and implementation. Commit co-signing required manual opt-out. Control metrics deteriorated: the author estimated comprehension below 20 percent on production merges.

Multiple developers interviewed in the post described identical patterns. Whole Jira backlogs routed through ACLI-connected agents yielded convincing diffs without architectural review. Context switching filled the 30-minute windows previously spent on implementation. Peer code review became untenable once the submitter could not answer basic questions about changes.

Operational consequence is measurable loss of system ownership. Organizations shipping via agent swarms now face elevated incident rates traceable to unexamined diffs. The post's one-month abstinence experiment functions as a controlled reversal: baseline comprehension and review latency returned within two weeks.

Future tooling will embed mandatory comprehension checkpoints. Repositories requiring signed human understanding attestations before merge will separate agent output from production paths.