Harness Engineering
6 articles
Why AI Agents Need a Harness, Not Just a Prompt
Prompt engineering was 2023. Context engineering was 2024-2025. The next phase is Harness Engineering, the runtime environment that makes AI agents reliable, verifiable, and controllable in production.
We Built a Harness Before It Had a Name
Before [Faros](https://www.faros.ai/blog/harness-engineering "Faros AI's public writeup naming the five-layer harness structure"), before [arXiv](https://arxiv.org/abs/2605.13357 "The 2026 paper that first formalized Harness Engineering as an academic discipline"), before [Strapi](https://www.ifttd.io/episodes/strapi "Strapi CEO Pierre Burgy discussing harness engineering as Strapi's next technical bet, July 2026"), there was a .claude/wiki/ directory. Four projects, four generations of evolution, one conclusion. The most effective harness for AI agents is the one you build from your own mistakes.
An AI Coding Harness, from Three Tmux Panes to Ten Agents
The first version of a multi-agent harness ran three personas in tmux panes, one of them Claude Code itself. It worked, right up until Claude tried to drive the thing headless and hit a wall no amount of retrying could get past. Here's the wall, and the ten-persona pipeline that replaced it.
A Coding Agent Breaks Its Own Rules Before Its Context Fills
The common explanation for an agent that suddenly ignores its own rules is a full context window. That's real, but it's not the only mechanism, and often not the one actually firing. A separate, measurable failure shows up under pressure, independent of how much context is loaded.
Our AI Agent Read the Rule and Took the Shortcut Anyway
A harness told an agent the correct command was in Taskfile.yml, and forbade the shortcut. The agent read the rule, found the correct command, and took the shortcut anyway. The fix wasn't a clearer rule. It was removing the shortcut.
Spec Kit, BMAD and OpenSpec, and the Cost of Planning First
Three tools claim to do the same thing, turning a request into a spec before any code gets written. They don't agree on how much that should cost, and a controlled benchmark says the cost is real. What their own review mechanisms reveal, and what we kept from two of them.