GitHub Copilot is the only AI we’re allowed to use where I work.
I don’t use it heavily. I had to turn off the auto-suggestions, because they were almost always wrong, and the hijack the tab key, which makes writing my hand painful (having to hit escape before every tab). I tried looking for a way to change the keybinding a year ago, but it didn’t seem possible.
I will occasionally use the Ask sidebar to see if it gives me some good options or for a quick syntax lookup. That’s about it.
It got Agent mode recently, which it defaults to. When this first happened, I went to ask a question and the agent just started doing stuff and destroyed my code in a way that undo didn’t fix. I had to go back to the commit history and make some manual changes to get back to where I was. I never used the agent again.
There has been some talk about getting Claude Code, but the problem is that it takes so long to make these enterprise deals and get something rolled out, that there is something new by the time we get the last new thing deployed. The pace of evolution in this space is faster than the enterprise moves.
I can only assume that when the next model releases, people will say the current Claude Code isn’t very good… so I’m probably not missing much in real terms. It kind of reminds me of video game graphics. The bleeding edge of the current-gen is always talked about as if its photo realistic and so amazing… then looking back at the same game 10 years later, and it looks like trash compared to the new bleeding edge. So maybe it’s not worth getting so worked up.
Yeah, this is exactly why you see all these surveys saying that AI hasn't given any increase in productivity. Especially when GitHub Copilot is still the #1 AI coding tool.
People were sold a spoon, get a spork, and end up eating with their hands.
There's a layer between your input and the model called an agent harness. It's the bit that guides the model how to traverse the file system, where to search, how the codebase is architected, how to navigate the monorepo.
When you say "Add a default $5 tip to the dialog screen titled 'Tip this waiter?'", what the harness does is supply information on where the strings are, the dialogs, and where the design style containing the PrimaryButton might be.
Cursor is excellent at this and probably pioneered the whole approach. Copilot hasn't really bothered to be more than a wrapper.
The models are the same. The agent implementation is different. I can confirm Claude Code performs much better than GH Copilot with the same Claude models.
I don’t use it heavily. I had to turn off the auto-suggestions, because they were almost always wrong, and the hijack the tab key, which makes writing my hand painful (having to hit escape before every tab). I tried looking for a way to change the keybinding a year ago, but it didn’t seem possible.
I will occasionally use the Ask sidebar to see if it gives me some good options or for a quick syntax lookup. That’s about it.
It got Agent mode recently, which it defaults to. When this first happened, I went to ask a question and the agent just started doing stuff and destroyed my code in a way that undo didn’t fix. I had to go back to the commit history and make some manual changes to get back to where I was. I never used the agent again.
There has been some talk about getting Claude Code, but the problem is that it takes so long to make these enterprise deals and get something rolled out, that there is something new by the time we get the last new thing deployed. The pace of evolution in this space is faster than the enterprise moves.
I can only assume that when the next model releases, people will say the current Claude Code isn’t very good… so I’m probably not missing much in real terms. It kind of reminds me of video game graphics. The bleeding edge of the current-gen is always talked about as if its photo realistic and so amazing… then looking back at the same game 10 years later, and it looks like trash compared to the new bleeding edge. So maybe it’s not worth getting so worked up.