Won't this be bad for agents since now you need to provide this new API in their prompt, as opposed to just regular git which the model has seen enough in its training data
replyhistory suggests so .... people do keep trying to make agent native tools and workflows, but time and time again it turns out to be better just to expose raw inputs and tools to them and let them work with those. See skills beating MCP in most cases where their purpose overlaps for example - it's more effective just to let an agent write git commands than give it a "git tool" with a structured interface. People don't seem to grok the intuition of how heavily biased training on trillions of token of human language and existing software code makes the models towards working well with raw input.
replysteve yegee was saying this in december, that he had multiple people at companies with heavy AI usage telling him that git cant handle 1000+ commits per hour that these companies are producing now, that the agent can't push because between the push command initiation and its finish another push happened
reply