27 March 2026
AI + Human = Better Than Either Alone
How harness engineering beat state-of-the-art on an NLP competition.
I competed in SemEval 2022 Task 4: detecting patronising and condescending language directed at vulnerable communities. 340 students. I ranked 10th manually writing detection systems and tuning models.
Then I built an AI agent system to do the same task. No human involvement. One shot, no prompt engineering. It ranked 22nd. Better than me, but not as good as my best manual work.
But when I added myself back in, directing the agents, refining prompts, verifying outputs, the system ranked 1st on the leaderboard.
The Numbers
- Me alone: 10/340
- My system alone: 22/340
- Me + system: 1/340
What This Means
The AI agent system I built is powerful. Left to its own devices, it’s above-average but not exceptional. But the system + human judgment is unbeatable. The AI does the heavy lifting. I add direction and verification.
This is harness engineering: designing the right constraints, feedback loops, and human-in-the-loop architecture so that AI systems do reliable work.
When I started building the system, I didn’t call it that. OpenAI published that term later. But I was solving the same problems: how do you keep agents on track? How do you verify they haven’t hallucinated? How do you build a system that stays reliable?
I built guardrails, prompt injection protection, agent logging, and orchestration patterns from first principles. The system needs structure. Humans need to stay in the loop.
The system alone is useful. System + human is unbeatable.