An essay on the GitHub blog from 27 July says something simple and rarely said out loud: the biggest jump in working with AI agents comes from how well you master the shell around the model - the harness. The model is the smaller part of the equation. A thesis that matches the everyday experience of anyone building with agents.
- GitHub published 'The harness is all you need (mostly)' - Burke Holland's take on working with AI agents.
- The thesis: the gain is in understanding the harness - the tools, rules and subagents around the model; new extensions and configurations are mostly 'tricks.'
- An essay with examples from his actual work, not a study with numbers - and that's exactly why it's honest reading for beginners.
First the word, because it isn't clear to everyone. 'Harness' literally means a set of straps. In the world of AI agents, that's what everything around the model is called - the tools it has access to, the rules it follows, the subagents it sends tasks to, the memory you give it. The model is the engine. The harness is the car.
I've built systems before AI was even a word in them - processes with nothing clever in them, just order: who does what, when, and by which rules. That's exactly what the harness is, just with a model instead of a person in the middle. The same model in two different harnesses gives radically different results: with clear rules, gates and checks it does the job; cut loose and bare, it flounders. The difference isn't in the intelligence of the engine. The difference is in the road you built for it.
For you, if you're just stepping into this kitchen, the takeaway is freeing. You don't need to chase every weekly novelty and every new plugin. Pick one tool and learn it all the way through - what it can do, where it messes up, how it's harnessed. Speed lives there. The catalog of novelties can wait.