Repository Harness, Part 2: 30 Runs Later, It Wasn’t Cheaper. It Was More Predictable.

Short summary

The author benchmarks a 'Repository Harness' approach against monolithic AGENTS.md files for coding agents, running 30 measured executions on a real Godot/C# game project. The Harness uses a manifest-based routing system to selectively load only relevant documentation per task. Results show the approach didn't dramatically reduce token usage but made agent behavior substantially more predictable by loading only task-relevant context.

• •Repository Harness routes context selectively via harness.yaml manifest instead of loading entire AGENTS.md
• •30-run benchmark on a Godot/C# game project shows improved predictability over token savings
• •Key fix: replacing broad load_when labels with strict routing rules (directory scopes, context budgets, escalation rules)

Generated with AI, which can make mistakes.