Member-only story
Six Agent Harness Techniques That Actually Worked for Me
What building production AI agents taught me about context engineering, routing, cost optimisation, human intervention, and evaluation.
13 min read2 days ago
Non members can read for free here
Building an AI agent has become relatively easy.
With frameworks like LangChain and LangGraph, we can give an LLM access to a few tools, let it decide what to call, and build a working prototype in a fairly short time.
But getting that same agent to work reliably in production is a different story.
While working on agentic applications, I noticed that many of the problems we encountered weren’t necessarily because the underlying model wasn’t capable enough.
- The agent could understand the user’s request.
- It could identify the appropriate tools.
- And in most cases, it could produce a reasonable response.
But there were other problems.
- We were spending unnecessary tokens passing information the agent didn’t need.
- Some queries were getting routed to the wrong workflow.
- The agent would sometimes make assumptions when important information was missing.
- And when something went wrong, it wasn’t always easy to understand which part of the workflow had caused the…
