KakaoPay FE Harness: Model Improvements Reduce Agents from 21 to 11

• Harness composition was reduced from 21 Agents and 29 Skills to 11 Agents and 17 Skills due to improved model performance
• Mechanisms to compensate for model limitations, such as step splitting, strong instructions, and keyword routing, were removed as they became unnecessary or increased token costs
• Team-specific convention documents, execution-result-based verification, and release gates remain necessary due to the model's tendency to evaluate its own results favorably
• Verification through observing actual behavior is essential even for enforcement mechanisms; without quantitative metrics, the harness accumulates indefinitely

Harness Composition and Enforcement Methods

• Path-specific rule files are generated and Hard/Soft rules are applied through domain knowledge presets and initialization Skills
• Seven Hook scripts are used to enforce delegation, prevent step skipping, block mid-execution termination, and record verification results
• Three types of Reviewer Agents (Maintenance, Convention, Accuracy) review in parallel, with final confirmation based on actual failure conditions

Issues Encountered During Development

• Blocking failed due to a PreToolUse Hook output format error, which was resolved by reimplementing the correct format
• An issue occurred where lint failures were recorded as passes due to exit code errors when executing pipe commands; this was fixed to record pipe commands as failures
• Work interference between sessions caused by using git stash was addressed by adding a rule prohibiting the use of that command in Skill documentation