Addy Osmani

A decent model in an excellent harness beats an excellent model in a bad one. Agent Harness Engineering, published 19 April 2026, carries that claim as an equation credited to Viv Trivedy: agent equals model plus harness. Trivedy's team moved a coding agent from Top 30 to Top 5 on Terminal Bench 2.0 by changing only the harness. Osmani's displayed function is Member of Technical Staff at Anthropic; earlier pieces place him at Google and Google Chrome.

He puts the bottleneck at verification rather than at writing code. In Agentic Verification (14 October 2026) the consequence is operational: tests are how an agent perceives the world and knows it has finished, end-to-end tests are the oracle hardest to fool, and the core loop should run under a second. His closing instruction is to write the check before the code.

Cognitive Surrender (5 May 2026) locates the failure in operator posture rather than in the tool: borrowing the model's confidence as a substitute for personal understanding. Throughput metrics cannot separate understood work from rubber-stamped work, so his countermeasures are small and physical: PRs around 100 lines, regular solo keyboard time, adversarial prompting. A year earlier he refused to let vibe coding and AI-assisted engineering share a name: the conflation misrepresents the discipline. Of the three joys of engineering he describes, the joy of knowing is the one most exposed to erosion by agents, and he offers no heuristic that restores it.