Day - 1 of Harness Engineering
It's Not the Car, It's the Driver: the model isn't the problem, the environment you give it is
Day 1 of Harness Engineering, just dipped my toe in this pond and already got hammered by new terms and problems I didn't even know could be fixed. So, here's the explanation of what I read in my understanding.
Everyone using AI has gone through "I need a paid subscription this is just not cutting it..". But your model itself is not bad, you are just not providing it the right environment, It's like taking an exam without knowing the syllabus or study guides.
Your model is the car and the Harness is the driver. Car has the power of a great engine but still needs a skilled driver to give it direction, So does your Agent.
Your Agent isn't dumb - It just gets stuck at places, doesn't know any escape routes.
In 2025 - A team ran an experiment using one of the best models of the time Opus 4.5, The task was simple : "Build a 2D retro game editor"
Result :
- Without any Harness - The agent spent 20 mins and burnt $9 worth tokens and returned an application that didn't work and still had runtime errors.
- With Full Harness - The agent was provided with a Planner, Generator and evaluator. The agent spent 6 hours and used up $200 worth tokens but the output was a fully function Playable Game.
Where did the Agent get stuck :
Let's break this down to points.
- Vague Requirements : The agent only understand what you tell it. The more it has to guess itself the more chances it can mess up. Telling a driver to just get to your home with giving them direction leaves them guessing the route and you'll never reach your destination.
- Implicit conventions not provided : Your agent doesn't know the difference between your normal and his until told. Its like asking someone to fill the gas in your petrol car without telling them it's petrol, what if they fill it up with diesel?
- Incomplete Environment : Your agent have a complete environment before starting the tasks. A car without tires won't until you install them.
When Things Fail
When things fail, Don't swap the model at first instance. Check the gaps in your Harness and fix them.
The core concept of Harness Engineering is Execute -> Observe Failure -> Fix the broken harness layer -> re-execute..
Yes, there are layers of harness...
- Task specification : Pretty self explanatory. Its an AI agent not a mentalist, explain your task in detail and depth.
- Context specification : AI does not have your memory or vision, any decision made for towards the project must be specified to the Agent. The agent can not follow rules it doesn't even know exist.
- Execution environment : Don't hand a driver a half-built car. tools install errors burn the whole context window, that's tokens spent fixing your env, not project.
- Verification feedback : A race with no finish line marked? Where does it end. The agent wouldn't know when to say it's done.
A simple fix
One file - Agent.md the holy grail of any project.
Just by adding an Agent.md file with the Project Description, Architecture, and verification methods. A dumb agent becomes an agent that does the job.