Skip navigation
I thought I had an agent problem. Turns out, I had a harness problem.
0:00

I thought I had an agent problem. Turns out, I had a harness problem.

#harnessengineering #aiagents #agenticai

I thought I had an agent problem. Turns out, I had a harness problem. I recently hit a roadblock while experimenting with a multi-agent research workflow. One agent planned what needed to be investigated. Another gathered evidence. Another analyzed it and made the final decision. Each agent worked surprisingly well on its own. But there was a subtle problem: Researcher finished ≠ enough evidence to make a decision. A tool could fail. Evidence could be missing. Sources could conflict. Yet the next agent could still happily continue. My first instinct was to fix the prompts. But “research more if needed” or “only answer when you have enough evidence” pushes system reliability back onto the LLM. The better question was: what should the system around the model enforce? That led me deeper into harness engineering — designing the runtime around the model: context, state, tools, control flow, retries, validation, budgets, observability and stop conditions. The model can reason. The harness determines the conditions under which that reasoning becomes a reliable system. That distinction is becoming increasingly important as we move from single LLM calls to long-running agents. Follow @thatmltechie for more AI engineering. #HarnessEngineering #AIAgents #AgenticAI #AIEngineering #LLM MachineLearning GenAI SoftwareEngineering

Comments

Comments from the original post on Instagram

No comments yet