Pedro Domingos says the model swallows the harness, and Elon Musk replies that the same thing happened to Tesla FSD's C++ code
That didn’t take long: the model swallows the harness.
Context
Two prominent voices argue that the scaffolding around AI models tends to get absorbed into the models themselves, a pattern that matters for anyone building agents today.
An agent harness is the program around a language model that organizes its calls, tool use and information flow. Domingos, a University of Washington computer science professor and author of The Master Algorithm, linked a September 28 arXiv paper titled Harness Learning Enables Generalizable Test Time Adaptation. Tesla's own release notes for FSD v12 said the city streets stack became a single end to end neural network, replacing over 300,000 lines of explicit C++ code.
Deep dive
On October 11, Pedro Domingos posted: That didn't take long: the model swallows the harness. He linked a new paper from researchers including Ruslan Salakhutdinov and Andrea Zanette.
According to its abstract, the paper treats an agent as jointly defined by its model and its harness, and introduces harness learning, which trains a proposer model with reinforcement learning to revise a solver's harness using execution feedback. The authors frame harness revisions as playing the role of weight updates, and at test time the proposer refines the harness for a new task without changing any parameters. They report that the ability to adapt transfers to unseen reasoning and multi hop question answering tasks.
Strictly, the paper has a model learning to rewrite the harness rather than replacing it, so Domingos' line reads as a broader comment on where agent design is heading.
Elon Musk replied a few hours later: Same thing happened with Tesla FSD. The model ate more and more of the C++ code, until there was nothing left. That matches Tesla's description of FSD v12 in its 2024 release notes, which said a single end to end network trained on millions of video clips replaced more than 300,000 lines of explicit C++ code.
What to expect next
Watch whether harness learning results replicate on harder agent benchmarks such as coding and long research tasks, and whether major labs start shipping models that write or tune their own scaffolding.
Also watch how this shapes the agent tooling market, since startups that sell orchestration layers could see more of their value absorbed by the models, as Musk says happened to Tesla's hand written driving code.