What happened to tab autocomplete?
9 points by iacore
9 points by iacore
Does anyone still use them? Now that agentic programming software are available.
not sure if you mean LSP or copilot-style autocomplete, but I still use the LSP autocomplete. It's deterministic, reliable and useful. If you mean copilot-style autocomplete, I don't like it or use it. I tried supermaven in summer 2024/2025 for a couple of months, but ultimately found it more annoying than useful. What mostly discouraged me was the worry that I might develop The Copilot Pause
(I do use LLMs/agents/whatever they're called this week, but i usually have it work on some support tooling, sometimes migrations/refactors in the background while i continue working on whatever I was working on.)
I do not use LLMs. I use LSP and ctags autocomplete.
Still going strong! đź’Ş
I never even got on the LSP boat; I just use vim's ctrl+n mapped to a conditional tab. So I type a couple letters, hit tab, and it auto fills based on existing words in other open files. I like that it is consistent for all languages and usually obeys my command instead of trying to usurp my authority.
As a nvim user, I have to ask: what if the needed word is not in your opened files? I use onmi complete with lsp to get an overview of the shape of the object I'm working on.
More specifically, I can't remember (or I don't care) all the methods a framework has.
In that case, I'll just type it out from memory, or maybe copy/paste it from the docs or library source. I've literally never actually used any of those intelligent completion things but my second-hand experience, trying to help people in support channels, makes me think they're not very reliable anyway. People often come in and are like "i can't believe method foo doesn't exist" and im like..... it does, and it is even called foo. Did you try it? Then they say no, it wasn't in the intellisense list. so meh.
idk though i can see the benefit if they do work, as long as it never pops up on me mid -typing, that'd drive me nuts!
It's also very nice when working with the same names across multiple file types. e.g. Mentioning type names in markdown files within a source repo or completing ticket names in comments from the ticket text pulled in from the bug tracker.
Yes of course. Why would I stop using it when programming?
What happened to tab autocomplete?
Nothing! It isn't as if editors removed the functionality, it's still there and still useful.
Does anyone still use them?
Also yes! Literally minutes before I clicked this link!
I actually much prefer tab-autocomplete. Both the traditional in-editor parsing, and backed by LSPs. I will admit I haven't used a coding agent, but then, I haven't run into a situation where I thought it would help.
It doesn't hurt that I work for myself, and don't have layers of management forcing agentic coding tools down my throat to justify their expenditure.
Maybe I have my head on backwards about it, and I'm open to hearing about interesting use-cases. I'm kind of afraid there's some amazing benefit I'm missing out on.
Yes, absolutely. Good auto-complete is just a very useful tool. And it's totally worth the time to have it set up correctly on our code editors, IMO.
But honestly, i don't really understand this question. I am using gen-AI quite a bit at $work, and i don't see how it could replace autocomplete. There are many things that are easier/faster/more convenient to do directly ("by hand") than using gen-AI. Good code editing tooling is very important in those cases. And even when using mainly gen-AI to drive work, i personally find that i need to constantly review and edit the things the AI generates. Sometimes it's fine to ask the AI itself to make those changes, but sometimes it's just easier to them myself.
What prompts this question? To me it feels a bit like asking "now that LLMs are a thing, is anyone still using version control systems?"
i believe the question is meant to counterpose “traditional” LSP style autocomplete against generative, Copilot style, next-token(s)-prediction style that is now standard in many IDEs.
edit: although on rereading the OP, i am less confident… maybe it is putting generative, Copilot style tab completion against running an agent without an editor?
IDK either. But you reminded me of that "autocomplete from LLM" mode that was the way Copilot worked initially, where if you paused typing, the LLM would suggest something without prompt. I had forgotten about that! I remember trying it out and having to disable it because i found it utterly annoying. It's like someone talking to you while you're trying to think; please shut up!
It's likely that OP meant that way of using LLMs, which is the most comparable thing to autocomplete i guess.
I for one have no interest in LLM-powered completions that actually write code for you. Even if they work decently (a possibility that I concede out of apathy, but remain highly skeptical of), I'd be too worried about becoming dependent on them and my ability to write my own code atrophying as a result. Once all these LLM agentic coding services decide they can't hemorrhage money anymore and adjust their pricing models to actually cover costs, I don't want to be in a position where I have to spend a significant chunk of my income to keep them doing my job for me.
I really liked Cursor's autocomplete through late 2024 – mid 2025, but it seemed to get significantly worse as they shifted towards agents. The autocomplete would try to make bigger and bigger guesses and almost always get something wrong, and none of the settings that were supposed to tune it actually worked for me.
I haven't found anything else that seems as good as Cursor was back then, and I've shifted mostly to Claude Code at this point.
Pre-LLM autocomplete has always been better for me, except if you really want to "autocomplete" a whole function, which is still 50% wrong last I checked.
If referring to the gen-ai github copilot style autocomplete, i went back to this style of LLM assistance after i went back to hand writing code somewhat recently. Zed finally supports local models for this and for me it's slow enough that it doesn't pop up a suggestion unless im sitting there thinkng or legitimately stuck. Occasionally it saves me a few hundred keystrokes but normally the suggestion is ignored without any issue.
I use the "classic" tab-complete based on semantic analysis in the IDE all day every day. Constantly.
I tried the LLM-based AI tab complete for a while but the latency was too high to make sense. It was slow enough that either it would show up late and interrupt me, or I would break my own flow waiting for it. It felt like pair programming with two people sharing the same keyboard at the same time.