Stop tweaking prompts: engineer the harness around your production agents

The transfer. The argument: build production agents like coding agents. Instead of hand-wiring Python or TypeScript tools and guardrails, define instructions.mmd at the agent root, keep skills in a skills folder, and let a platform package it. The demo shows this in LangChain Managed Deep Agents, with an agent definition in about 30 lines of code. Markdown becomes the control surface for rules, workflows, memory, and guardrails. Changing behavior becomes editing documents, not redeploying services.

The feedback loop. Improvements come from two loops. First, the agent flags uncertainty during real user conversations and escalates to a team inbox. A human confirms or corrects the answer, and the lesson is stored as shared and per-user memory. In the demo, a question about Archon build work trees was escalated, corrected, and a later similar question from another user was answered with confidence instead of uncertainty. This is a practical pattern: production traces become training material without retraining a model.

The evidence. The numbers are modest: 14 real questions, starting at 4 correct and reaching 13 after lessons and trace-based fixes. That is a useful demo, not proof. There is no comparison to a hand-coded agent, no benchmark against another platform, and no cost or latency data. The video is sponsored by LangChain and the demo is the centerpiece, so read the claims as vendor evidence, not independent results.

The control tradeoff. The second loop uses LangSmith traces to propose changes to the agent's instructions.mmd, and a human approves before the fix is applied. That is the right governance model: the agent can suggest and apply after approval, but it cannot rewrite its own rules freely. Versioning and rollback lower the risk of a bad edit. The open question is how this scales when a team runs dozens of agents and the review queue becomes a bottleneck.

What's missing. The video does not address what happens when markdown harnesses grow, memory conflicts, or instructions bloat. It also does not compare the managed platform with an in-house pipeline. Teams should not assume a one-to-one mapping between coding agents and production agents always holds: customer-facing agents have different accountability, privacy constraints, and failure modes than coding assistants.

The takeaway. Harness engineering is continuous. The transferable discipline is to treat agent rules as versioned configuration, collect traces as first-class data, keep a human in the approval loop, and use lessons to reduce escalations over time. The specific product may be useful, but the actionable signal is the workflow, not the vendor demo.