Descriptions:
Tejas Kumar, AI developer advocate at IBM, delivers a methodical first-principles explanation of what agent harnesses are and why they are essential for building reliable AI systems. Speaking at the AI Engineer conference, Kumar addresses a term that has become widespread but inconsistently defined — distinguishing the machine learning meaning (a test suite for model outputs) from the AI engineering meaning: the complete infrastructure layer surrounding a model that anchors it to a stable, controllable environment.
Kumar breaks the harness down into its core components — tool registry, context management, agent run loop, and guardrails — and builds a working browser-use agent in Node.js live on stage to illustrate each layer. Without a harness, the agent fails silently: it clicks an upvote button, hits a login wall, and reports success anyway. Kumar then incrementally adds guardrails (max iterations, context compression thresholds) to show how each layer transforms unreliable behavior into predictable output. Claude Code, Cursor, and Codex are used as concrete real-world examples of fully harnessed agents throughout the talk.
The central argument is about controlling the uncontrollable: since the underlying model is a black box that could theoretically serve a different model tier without notice, the harness becomes the only stable surface an engineer can actually reason about and test. The talk is structured as a progressive build, making it practical for engineers at any experience level who want a concrete mental model for agent reliability.
📺 Source: AI Engineer · Published May 17, 2026
🏷️ Format: Deep Dive
