When people ask me why I like working in AI, I usually have a few stock answers.
For the techno-utopians, I want to change the world and contribute to the tremendous impact AI will have on humanity (only the positive one though 😇). For hiring managers, it's about using AI to build real production-ready apps. For the cynics (or are they realists?): so I still have a job in a few years.
The real reason though is so I can watch researchers come up with increasingly complicated terms to describe increasingly simple concepts.
Amidst all the jargon, one word that I think will stand the test of time is "Harness".
I see it used in different contexts: an eval harness, a test harness, an agent harness.
Language models have tremendous potential, it is our job as builders to harness this into something useful. We do this by dipping into a few age old lessons that good leaders know:
Context creates alignment. The best leaders are able to help align their teams on outcomes, providing clarity about the context while preventing cognitive overload, letting their teams figure out the how. For agents, we use prompts to clarify outcomes, context engineering for the right amount of information and a set of tools to let them figure out the how.
Feedback matters. Teams can figure out the best solution only if they have the right feedback loops in place. For product teams, this might mean the right tools for product discovery. For engineers, robust testing and observability. For agents too, these feedback mechanisms are similar: evals, telemetry, linters, observability and analytics. All guided by a well defined outcome.
Constraints unlock creativity. Good leaders know how to create a framework to channel their team's creativity, good designers often use constraints to overcome blank canvas paralysis. Both lead their teams from experimentation to results. Agents benefit from well designed prompts, planning tools and delegation to sub-agents. These create a system that guides an agent to results, ensuring that it doesn't hallucinate intent and stays on-track.
Collectively, this creates a harness that keeps the model focused on the outcome, giving it enough tools and context to achieve and verify the outcome. We transform it from a token predictor into an outcome machine.
The parallels and lessons are often uncanny, and potentially dangerous. Which leads me to the final reason I work in AI, the one I don't tell the utopians or the cynics. I refuse to accept as inevitable the use of these analogies to stop investing in the people around us.