Coding Agents Can't Verify Their Own Work — Here's Why That's a Problem

Coding agents typically report task completion in natural language, but this self-assessment is unreliable since the same agent that writes code also judges whether it is finished. Common failure modes include misread test output, incomplete verification commands, and assumed successes that were never confirmed. A developer building an open-source Go runtime called VichuFlow argues that verification must be handled by a separate, independent authority — much like CI systems in pull request workflows. The proposed architecture uses a persistent state machine to gate workflow transitions based on observable evidence such as exit codes from test, lint, and typecheck commands, rather than agent declarations. VichuFlow is an early-stage project designed to coordinate coding agents within structured, durable workflows where the runtime — not the agent — decides when a task is truly complete.