Addy Osmani
Author of X Articles on agent harness engineering, cognitive surrender, and loop engineering; frames agent work as model plus scaffolding plus human judgment.
Addy Osmani is the author of the X Article "Agent Harness Engineering," which argues that the practical performance of coding agents depends heavily on the harness around the model: prompts, tools, context, hooks, sandboxes, subagents, memory, and feedback loops.1
The article connects public discussion by Varun Trivedy, HumanLayer, Anthropic engineering, Birgitta Böckeler, Fareed Khan, and Fred K. Schott into a broader framing: the industry is converging on harness patterns that can make a decent model with a strong harness outperform a stronger model with weak scaffolding.1
Osmani also wrote "Cognitive Surrender," distinguishing productive cognitive offloading from the risk that AI output becomes accepted as the engineer's own view without independent understanding. That article connects AI coding practice to Cognitive Surrender, comprehension debt, and review/verification habits that keep agents in a mutual-amplification role.2
Osmani's "Loop Engineering" article extends that same agent-systems view from a single-agent harness to recurring orchestration. He argues that engineers are moving from prompting agents directly toward designing loops that discover work, dispatch agents, verify outputs, preserve durable memory, and decide the next action while the human keeps responsibility for review, understanding, and quality.3
Mentioned in 4 pages
Engineering with agents
Thinking next to AI