Daniel Miessler, the security researcher who writes the newsletter Unsupervised Learning, published an essay on July 29 answering a question that venture investor Martin Casado had posed publicly. Casado, a general partner at Andreessen Horowitz, said he cycled between three contradictory beliefs about AI agent harnesses: that less scaffolding is better because the model itself is what matters, that model providers will eventually train the harness into the model, or that a harness carries value the underlying model can never absorb. Casado said he could not tell which belief was correct.
Miessler’s answer is that all three beliefs are partially true because a harness is not one thing. He splits it into two components that age in opposite directions. One half is instructions for how to execute a task, and he says that half loses value as models get smarter, an argument that echoes Richard Sutton’s Bitter Lesson: hand-tuned procedural guidance looks weaker next to a more capable model with every new generation. The other half is context about what the operator actually wants, who they are building for, and what a good result looks like. That half, Miessler argues, grows more valuable as models improve, because a smarter model can do more with the same intent once it receives it.
The distinction lets Miessler resolve Casado’s second belief too. Model labs can and will post-train execution ability into their systems, closing the gap on procedure. What they cannot train in is a specific operator’s context, constraints, and taste, because that information lives outside the model and shifts with every task. Miessler calls the practice of capturing and delivering that context Intent Engineering, and says it is the discipline that should survive even as harnesses otherwise shrink.
The framework lines up with evidence AI Insiders is also covering today: OpenAI found that adjusting two settings inside an agent’s harness roughly tripled its score on a coding benchmark, with no change to the underlying model. That is a direct data point for Miessler’s claim that the scaffolding around a model, not just the model, is doing measurable work.
For teams running agents in production, the practical read is a diagnostic one. When an agent underperforms, the fix is often the harness rather than a model swap. Auditing how clearly a system captures user intent before touching the model tier is the cheaper repair, and the one more likely to hold up across the next model upgrade.
Daniel Miessler published this analysis on his blog on July 29, 2026.
