0:01March 2026. Andrej Karpathy says he hasn’t typed a line of code since December.
0:06One of the most influential AI researchers in the world is no longer spending his time writing code. So what is he doing instead?
0:14He's designing the systems that let agents do the work. And there's a discipline behind it that most engineers are missing.
0:20Mitchell Hashimoto, co-founder of HashiCorp, calls it Harness Engineering.
0:25And it is the biggest shift in how software gets built since we moved to the cloud.
0:30For the last three years, we have all been waiting for the models to get smarter. We thought the model was the bottleneck.
0:36But we've crossed a threshold. The models are now capable enough to do the work.
0:40The bottleneck is no longer the model. It is our ability to build a reliable system around it.
0:46If you just take a raw, highly capable model and point it at a codebase, it will fail.
0:51It gets stuck in doom loops. Trying the same broken fix ten times in a row.
0:56It suffers from false completion, telling you a feature is done without actually running the tests to prove it.
1:01And it causes architectural drift. Making local fixes that violate your global design rules.
1:07A raw model is smart, but it is deeply unreliable. It is like giving your junior developer root access, and no code review.
1:16This brings us to the core insight of this entire shift. An agent is not just a model. An agent equals a Model, plus a Harness.
1:24The harness is the scaffolding that turns a capable but unreliable model into a reliable work engine.
1:29As the engineering team at LangChain put it recently, "If you are not the model, you are the harness."
1:37So what actually is it? It is everything that surrounds the model.
1:40It is the execution environment. The isolated sandbox where the agent runs safely.
1:45It is the context layer. The maps and documentation the agent reads at startup so it knows where it is.
1:51It is the architectural constraints. The deterministic linters that block bad code before it merges.
1:57It is the orchestration logic. The workflow that tells the agent what to do next.
2:01And it is the verification layer. The system that forces the agent to test its own work before it claims to be finished.
2:09If you think this is theoretical, look at what the top teams are actually doing.
2:13OpenAI recently revealed they have built a production codebase of over one million lines of code.
2:18Exactly zero lines were written manually. A team of just three engineers over five months.
2:24Fifteen hundred pull requests merged, agents running unattended around the clock.
2:28Those engineers did not write the code. They wrote the system that wrote the code. That system is the harness.
2:36Stripe is doing the exact same thing.
2:38They are currently merging over thirteen hundred pull requests every single week written entirely by their internal agents, which they call Minions.
2:46These are fully unattended.
2:48A human engineer tags a bot in a Slack thread, walks away, and returns later to a finished, tested, and verified pull request.
2:55So let this sink in again, Thirteen hundred pull requests. Every single week. Written entirely by their internal agents
3:04That is only possible because Stripe spent years building a world-class execution harness.
3:11We have just seen examples that indicate a significant shift is underway.
3:15The top experts like Karpathy. The top companies like OpenAI and Stripe. They are no longer writing code. They are designing systems that write it.
3:25Reliably. At scale.
3:27That, is harness engineering.
3:30And that is where you want to be.
3:33So, how do you actually build an effective harness?
3:36In Video 2, we open the hood.
3:38The three technical layers every harness needs, the execution environment, the context constraints, and the orchestration loops, and how Stripe, OpenAI, and LangChain built theirs.
3:50I'll see you there.

Harness Engineering: The Bottleneck Moved

Created by Abhyuday DesaiUpdated 2 Oct, v2Runs 3 min 57 sec

0 views

About this video

The biggest shift in software engineering since the cloud. AI agents are now capable enough to write production code, but only if you build the right system around them. This video breaks down harness engineering, the discipline that's replacing traditional coding at the industry's top teams. Part 1 of a 2-part series.

What changed

v202 Oct 2026abhydesai3 min 57 sec

Re-rendered in the studio with chapters and a transcript. The narration is now by Alex, the series' voice; v1 was narrated by Mark.

v1 has no transcript to compare.

0 views0 likes on this version

v107 Apr 2026abhydesai3 min 59 sec

First published cut.

0 views0 likes on this version

Watch v1