GitHub Copilot: The Agent Awakens
111 comments
·February 6, 2025diggan
suyash
The irony doesn't end there, later in the post they say
"We’re excited to share a first look at our autonomous SWE agent and how we envision these types of agents will fit into the GitHub user experience. When the product we are building under the codename Project Padawan ships later this year, it will allow you to directly assign issues to GitHub Copilot, using any of the GitHub clients, and have it produce fully tested pull requests."
- Effectively they have completely automated SWE job, pair programmer was just a marketing speak, real intention is clear.
phil917
"Project Padawan" looks fairly similar to Devin, at least from a user experience perspective. From personal experience, Devin was pretty terrible so we'll see if Microsoft does any better...
nozzlegear
> it will allow you to directly assign issues to GitHub Copilot, using any of the GitHub clients, and have it produce fully tested pull requests.
I hope I can turn this feature off, i.e. that it's not a feature other users can use on my repositories. I'm already getting AI slop comments suggesting unhelpful fixes on my open source projects, I don't need "the anointed one" sending over slop as well – while replacing the work of real humans, to boot.
throw83288
> Effectively they have completely automated SWE job, pair programmer was just a marketing speak, real intention is clear.
Frankly, this was obvious to me since the Copilot Workspace announcement.
It's so hard not for me to not slide completely into nihilistic despair as a student right now. I chose wrong, and now there is nothing I can do. Day in and day out I'm talking about my projects and internships as if my entire field that I've dreamed about for the past decade isn't about to get torched. With the pace that this field is getting solved I probably won't even have enough time to "pivot" to anything before they also get solved or upturned as well.
Call me a doomsday prepper, but frankly I haven't heard a compelling argument against this line of thinking that is actually in line with the absurd development curve I've seen. 4 years ago these models weren't capable of stringing together a TODO app.
mywittyname
I've been in the industry long enough to have been around for a few crashes. My outlook is: this industry has always faced threats that looked like it was going to spell the end of our careers, but we always come out the other side better than ever.
I don't think LLMs are fundamentally more threatening than off shore developers were. Sure, we lost jobs during that time, but businesses realized eventually that the productivity was low and they wanted low level people who were responsible.
I think that will continue. We'll all learn to control these agents and review their code, but ultimately, someone needs to be responsible for these agents, reviewing what they produce and fixing any shitshows they produce.
I won't rule out the possibility of LLMs that are so good that they can replicate just about any app in existence in minutes. But there's still value in having workers manage infrastructure, data, etc.
konart
>Call me a doomsday prepper
>Seems obvious that I missed the boat on LLMs.
Don't worry. As the other commenter said it: we've seen it all a few times already.
I clearly remember how some people reacted to Ruby on Rails as if it's going to replace them just because it provides... well a framework.
LLMs won't replace even a junior dev anytime soon. Not to mention senior dev etc.
People who's main job was creating landing pages and simple shops might be in trouble.
toprerules
This must be your first hype cycle then. Most of us who are senior+ have been through these cycles before. There's always a 10% gap that makes it impossible to fully close the gap between needing a programmer and a machine doing the work. Nothing about the current evolution of LLMs suggests that they are close to solving this. The current messaging is basically, look how far we got this time, we will for sure reach AGI or full replaceability by throwing X more dollars at the problem.
semi-extrinsic
> Any "normal" engineering field will be solved with the right domain knowledge.
Oh no, don't worry. Nobody will be trusting GenAI only for Real Work like aerodynamics, structural mechanics, electromagnetics, plasma physics, you name it. For sure there will be (already is) AI-based surrogate models, fast preconditioners, superresolution methods etc. But you will for the duration of our lifetime need humans who understand both physics and programming to actually use the damn tools and ensure they are used correctly.
nickstinemates
If you want to thrive in this world you need to change your attitude ASAP. New tech waves happen all of the time. Embracing them is the path.
nonethewiser
I dont think making Copilot better by handling its bad ouput means replacing developers. And GitHub certainly isn’t saying the goal is to replace developers.
I accept that is how you are interpreting it and I can see the argument. But Github isnt trying to get one over in their messaging.
And besides I just dont agree with the idea that it takes the developer out of the loop. Whose controlling this better version of Copilot? Whose goals is it advancing? The developer.
jacobsenscott
The goal has always been to eliminate programmers.
Nobody wants to pay a bunch of desk workers six figures to make their business go brr, but they currently they have no choice. Trust me, every executive resents this to their core and they want all the programmers to go away - including github executives.
20 years ago you would hire a few expensive architects who would try and design the product in so much detail cheap jr programmers could build it. It didn't go well.
4GL languages tried to abstract away all the hard stuff - again it didn't go well.
"Low code" was big just before the AI thing. It didn't go well.
Attempts are outsourcing are constant.
Now we have LLMs. So far this has come the closest to the dream of eliminating expensive programmers. We'll see how it goes.
diggan
> I dont think making Copilot better by handling its bad ouput means replacing developers
The blog post goes through more than what I mentioned in my comment. For example:
> When the product we are building under the codename Project Padawan ships later this year, it will allow you to directly assign issues to GitHub Copilot, using any of the GitHub clients, and have it produce fully tested pull requests. Once a task is finished, Copilot will assign human reviewers to the PR, and work to resolve feedback they add.
How is that not trying to replace at least a small section of junior/boilerplate developers?
The developer might be the one who listens to the product team, maybe even creates the issue and finally reviews the code before it gets merged. But I'm having a hard time imagining the flow above as "Pair programming" or a developer working with a "co-pilot", as they're trying to say it's all about.
chipgap98
I don't think those are contradictory. The agent is iterating on the code it writes in order to be more useful.
diggan
> When the product we are building under the codename Project Padawan ships later this year, it will allow you to directly assign issues to GitHub Copilot, using any of the GitHub clients, and have it produce fully tested pull requests. Once a task is finished, Copilot will assign human reviewers to the PR, and work to resolve feedback they add.
This makes it seem like a very basic/rudimentary developer could be replaced by the new autonomous agent. Or am I misunderstanding what they're announcing here?
someothherguyy
What else would the goal of agentic generative AI be other than to replace humans doing the same work?
diggan
For me that's a no-brainer also. I guess my question is more about why their messaging is so messy and direction-less. Why not just say outright what their plan is?
whilenot-dev
Backlash towards the direction off of GitHub? I mean GitHub has to be one of the biggest assets when it comes to training data for source code, no? Projects should have already moved off of it when Microsoft got in the drivers seat.
827a
IMO: Copilot (and Devin) has always missed the mark. Its always been a lazy, bad product that feels like it was made by a team that doesn't even want to use it themselves. Going more agentic is only going to make it worse; Copilot's product leadership appears obsessed with more comprehensive replacement of their customer workflows, but ideal customers want deeper and more fluid integration into the workflows.
Its one of those facts that seems so obvious once you realize it, but no one at Github clearly does. Who is buying and using these things? Seriously, anyone at Github, answer that, not who you think is buying them, but who is actually buying them. The answer isn't CEOs or CTOs; its Software Engineers (or CTOs, for their Software Engineers). Github's leadership needs the answer to be CEOs or CTOs, because the scale of investment (to produce such a shit product) is so large that only per-customer revenue commensurate to the replacement of SE salaries justifies it.
I know of four companies (including my own) that had a corporate Copilot subscription for their devs, and over the past quarter/this quarter are replacing it with a Cursor subscription, at the request of their devs. I'm super bullish on Cursor and Supermaven.
- I think they understand their ICP way better.
- I think their ICP is actually excited to spend money with them.
- I think these new companies have demonstrated that they are more willing to build more than just a panel in VSCode; whereas Github is bogged down by legacy interests.
- I think this deeper level of integration into existing workflows is what pushes AI past the hump of "oh i want to use that". Speeding up existing workflows by 30% feels insanely good. It grows the pie. Smaller & smarter, not larger & derivative.
- I think, from a business perspective, MS/Github has and continues to royally screw up by literally subsidizing the cost basis of their competitors by building VSCode and hosting billions of lines of open source code competitor models train on. I love it as a user. But it costs them millions of dollars, and every dollar of that spend makes their competitors stronger.
WithinReason
I don't see a contradiction
rvz
This sort of obfuscation of Microsoft/GitHub's real intentions *IS* deliberate. Unfortunately it's not just them, but pervasive across nearly all AI companies.
> Are they maybe afraid of a potential backslash if it became more clear that they're now looking to replace (some) developers / the boring parts of development?
Did they care about the writers, musicians, artists, journalists that had their jobs displaced or currently reduced? I don't think so and they got away with it.
They don't care and in 2025, programmers of all ranks are next.
Look at their actions and don't fall for the blog posts or statements.
Tarucho
They are not going to say "we are making an agent to replace programmers. We will use their code and their guidance to train it"
lolinder
They'd better get on the IntelliJ integration fast— if I'm going to switch editors in order to use an LLM coding assistant, I may as well just switch to Cursor, which has a strong head start on them on the assistant and also has a much better autocomplete.
I'm honestly surprised to see no mention here of them moving to replicate Cursor's autocomplete—IMO that is where Cursor's real edge lies. Anyone can wrap a chatbot in some UI that allows you to avoid pasting by hand, but Cursor's ability to guess my next move—even several lines away—makes it far and away more efficient than alternatives.
twistedgost
The feature you are referring to was also announced in VSCode and is called Next edit suggestions for Copilot. Currently in preview: https://github.blog/changelog/2025-02-06-next-edit-suggestio...
petercooper
JetBrains has got its own version in the pipeline as well: https://blog.jetbrains.com/junie/2025/01/meet-junie-your-cod...
hassleblad23
Github has completely abandoned the Intellij Copilot plugin it seems. Even model selection is not supported. This is good for Jetbrains though because they have their own competing AI service. Jetbrains AI doesn't support multiline edits in tab completion or chat, but it does in the inline prompt mode (although its limited to the same file only).
armchairhacker
IntelliJ with Cursor-like autocomplete or Cursor with IntelliJ-quality general IDE tooling (lookup/rename symbol, diagnostics, and general UI) would be the ultimate editor.
IntelliJ’s autocomplete was really bad last time I tried it, and if it’s still only single line it’s still bad. Fortunately GitHub copilot in IntelliJ is good, maybe as good as Cursor except that it can’t delete/rewrite code or jump to different locations.
IMO agents aren’t nearly as important for either team to focus on, because they can be used outside of the IDE or in a separate IDE. I think the teams who develop the best agents will be big model-trainers and/or teams dedicated to agents, not teams writing IDEs.
jghn
Yeah. I just wish that VSCode didn't feel so crude coming from 10+ years using JetBrains IDEs. Things I feel are table stakes like nice test run/debug functionality seem like big hurdles. Perhaps it's just a learning curve & I need to get used to it, but whenever I dive into how to replicate functionality I feel is important it seems the answer is at best "it's complicated".
It's a shame as this is by far not the only thing in which I have interest that seems to have fully shifted over to VSCode
zn44
i've started using aider with https://aider.chat/docs/usage/watch.html works great and you can keep using jetbrains IDEs
theasisa
I tried Cursor a couple of years ago and wasn't impressed - has it improved a lot? I only use autocomplete, not the chat function and at the time found CoPilot superior.
baq
It has improved but you're missing out if you aren't using the big ticket features. I tab myself to solutions, too, but if there's a react view to do, I dish out the composer and am literally 10x faster - what would previously take a day now takes an hour. If there's an interface to create out of a json blob, I paste the blob and just tell it to make an interface, then clean up the types a bit, etc.
nirava
I'd written off AI autocomplete as pointless after trying GitHub Copilot's a year ago.
But Cursor's tab-autocomplete is actually really useful. It feels like it very much knows what I'm up to.
sturza
Years?
827a
Cursor is ten times better than VSCode and Copilot. Its extraordinarily good at reducing two-minute tasks to 10-seconds, and the more you use it the better you get at identifying these two-minute tasks.
Example (web dev): hit cmd+k --> "this is a two column layout. make sure the columns are the same size". It just does it. To do that myself I would have had to switch to a browser, google flex box, go to that classic flexbox cheat sheet that we all know and love, tweak around with the different values of justify-content and justify-self, realize that was the wrong approach, then arrive at the correct answer of making sure each column flex-grows identically. two minute task, now 10 seconds.
hit cmd+k -> "flow these columns one-after-another on smaller screens" done. thirty second task, now 10 seconds.
hit cmd+k -> "enable or disable the rendering of this component via props" done. new prop added, prop is flowed through to a `display` css property, easy.
The autocomplete is pretty good, but can get annoying. You definitely have to get used to it. However, the cmd+k quick fix thing is insane. Its literally made me at least 200% more productive, and I think that might grow to 300% as I learn to use it and it gets smarter (they just added Gemini 2.0 Flash; can't wait to try that out).
lolinder
I tried it last month on a medium size personal project and was blown away by the autocomplete. I'd previously staunchly refused to try it on the grounds that I'm too productive in IntelliJ, but at this point I'm most likely going to start paying for both.
I don't know if I'm ready to use it as a daily driver, but there are certain kinds of tasks—especially large refactors—where its ability to rapidly suggest and accurately make the changes across a file is incredibly valuable. It somehow manages to do all of that without ever breaking my sense of flow, which is more than I can say for Copilot's suggestions.
And yeah, I'm with you that autocomplete is the way to go. I think chat is a red herring that will have long-term negative effects if it's used extensively in a codebase. Autocomplete keeps you in touch with the code while still benefiting from the co-pilot, and Cursor's UX for that is far and away the best I've seen.
cruffle_duffle
I wish cursor was an extension of VSCode and not a fork.
diggan
Does is matter in practice? Is there stuff you can do in VSCode that isn't possible in Cursor? I'm not a user of either, so honest question.
gdhkgdhkvff
For one, you can’t debug c# code in cursor without using a hacky third party extension. Because the c# debugger is only licensed to run in official vscode instances. And only way you find out is you try to run c# and get a runtime error saying that it can’t run for that reason, you google/chatgpt the issue, find your way to some old GitHub Issues threads where someone mentioned that’s a possible solution.
machine_ghost
I don't know Cursor, but VS Code is a very full-featured editor with many years behind it; I rather doubt an upstart editor could achieve full feature parity with it so quickly.
But that's almost beside the point: even if it had perfectly identical functionality, people would still want to use VS Code, if only for its well-established ecosystem of extensions.
hidingOnBush
People are sleeping on codeium. I've found their AI assistant to be much better than cursor
thomasfromcdnjs
Cursor has obviously figured out marketing better.
I switched to Windsurf 2-3 months ago, feels a lot better for me.
satvikpendem
I was looking at Windsurf and Cursor as well, what are the differences?
astrodude
Agreed. Windsurf is a lot better
lolinder
How's their autocomplete? I'm honestly not interested in tighter integration of chatbots. What blew me away about Cursor was how much better it was at autocomplete. I honestly probably would have tried it sooner if people emphasized that strongly enough in online dialogs, but it weirdly always seems to get relegated to an afterthought compared to the flashy chatbot, which was... fine, I guess?
aresant
“Copilot puts the human at the center of the creative work that is software development. AI helps with the things you don’t want to do, so you have more time for the things you do.”
… until we train our model on your usage data and totally replace you
threatofrain
If we actually could get to a world where programmers can be replaced, we'll also likely find that vast swathes of the population will be replaced. Then we'll need a totally new conversation on how society should look. That conversation is coming no matter what, because there will not be a global consensus to stop ML development, esp. on the war front.
johnfn
In a world where programmers can be replaced, it's less clear to me that plumbers, janitors, electricians, construction laborers, etc, are also all out of a job.
suyash
Here is a saying that I really think summarises "AI will replace white-collar jobs and Robotics will replace blue-collar jobs".
reducesuffering
Think secondary effects. What does a world in which almost every programmer can be automated look like? It looks like massive extremely fast technological development, how to build and program robotics will be solved almost instantly. With solved robotics goes every other labor.
We don't get the same current X productivity with 1/100th the people. We get 100x productivity, controlled by a few people / megacorps, until they lose control of it too.
gokhan
As a programmer, I think it's easier to replace programmers than many other occupations. We work with generic DSLs called programming languages, which can be easily handled by AI. Most of what we produce can be easily parsed by an AI. We work in small, incremental tasks which is perfect for AI. Most of us use the same set of tools. What else? UI is similar everywhere, especially in LoB apps. What we produce is repetitive, most of us practically doing clever meshups of the code on the Internet.
Thank Crom I'm old. But I'm worried for my kids.
ericmcer
Doing the small incremental tasks is the easiest part of programming. All the complexity lies in taking giant specs and breaking it into manageable goals and then small tasks. AI is still pretty far from doing that, but I agree, I would not want to be a teenager right now and face the prospect of entering the workforce after 10 years more of progress.
Unless work is a thing of the past by then, and we were the last generation of suckers who lived under the reality where 40hrs/week was just the way life is.
machine_ghost
In the long run, yes, but if programming were so rote and simple, we all wouldn't have jobs in the first place.
Look, technology has always replaced jobs, throughout history (how many of you know professional lamp lighters, elevator doormen, or switchboard operators?) But the thing is, tech has always replaced the jobs that are easiest to automate ... not the "intellectual work".
I would be far more worried about being a real estate agent, food preparer, or heck, even a lawyer (for some specialties at least). I think they're all at more risk than programmers.
hirako2000
More time for the things I do?
All AI tools I've used turned my daily into fighting the agent.
While LLMs can save time on tedious repetitive tasks, they are atrocities in producing bug free code. Even if they reached 99.999% accuracy without, they wouldn't be worth it: If I wrote a function and it turns out buggy, I can, even months later, investigate and find the culprit.
If an LLM introduces a bug, I would rather scrap everything. That is, every piece of code the thing produced for a given project.
If my boss tells me my job is to provide a patch, I quit.
And, LLMs have proven so far that they can't produce innovative solutions.
amelius
If they really put the human at the center then they'd contribute back and publish the weights so that we can run this locally and build our own AIs on top, etc.
shaneofalltrad
If they don't, it will be Cursor, Trae, Cline, Roo Code or Goose? Seems this is coming this year in a big way regardless if Copilot does it or not. I'm guessing we all have to pivot how we work or get left behind?
rvz
They are not hiding it anymore and are testing as to how much they can get away with total job displacement. First artists, then musicians, then writers and now programmers.
Better prepare for 2030 then as I am already doing so [0].
But from this year and in the next 5 years I'd also read this very carefully [1].
[0] https://news.ycombinator.com/item?id=42563239
[1] https://www.weforum.org/publications/the-future-of-jobs-repo...
JohnnyMarcone
Care to share how you are preparing?
balls187
Was using Github copilot with VSCode. Found it really helpful for small things.
I gave Codium Windsurf a spin to evaluate it on an internal app built: a simple svelte-kit/typescript app ontop of Twilio. It's been on todo list to host it (on lightsail), so the suits can use it themselves.
Asked Windsurf to enhance some capabilities: adding a store library backing to dynamodb.
Windsurf's code failed to run. svelte-check reported 6 errors. I asked windsurf to fix the errors, and it did, creating 10 errors. Repeat one more time and resulted in 16 errors. If I wasn't busy, I would have seen how many errors it would get up to before I got bored. It felt like repeatedly opening a JPEG and resaving it.
Giving up on using Windsurf. Gonna try Cursor next. After that, back to Github.
dabinat
I use Copilot and it’s a game-changer for my productivity, but I really wish it was capable of natural language searching. So for example I could ask it “show me all places in the code where x is assigned a value but a flush() command is not immediately issued afterwards”.
simonw
I answer those kinds of questions by piping my entire codebase into a large context model (like Claude or o3-mini or Gemini) and prompting it directly.
Here's a recent example:
files-to-prompt datasette tests -e py -c | \
llm -m gemini-2.0-flash-exp -u \
'which of these files contain tests that indirectly exercise the label_column_for_table() function'
jgilias
I think Cursor can do this, if you @codebase, isn’t there something similar in copilot? E.g., your codebase being vectorized, indexed, and used as an embedding?
shaneofalltrad
I have had Cursor review all my file content solving similar things, but I would think it's limited to VSCode search capabilities and IMHO it's not great. I love how Pycharm handles indexing so search is fast and accurate. If they ever get agents going at the same quality as Cursor I would probably go to Pycharm for that advantage alone.
wrs
Cursor implemented their own code indexer so its RAG is not limited by VS Code search.
bramhaag
It can suggest terminal commands and ask you to execute them
conradfr
That's why you need Jetbrains local history feature.
ddalex
nuke a project or root is the best case scenario
snapcaster
The likely case is that it almost never does anything harmful. I've never once seen an LLM tell me to run rm -rf /
indymike
yet
acedTrex
I am so over this LLM obsession. Please let it end
magicmicah85
The autonomous agent stuff is what’s had me worried the past year as lots of open source projects have popped up that do similar capabilities. It’s really cool technology but it will replace humans, no matter how garbage the code quality is. I’ve seen garbage from humans, I’ve seen garbage from AI. As long as metrics are being met, business does not care how they achieve their goals.
I for one am going to welcome our new agent underlords as I’m all about self preservation.
ks2048
So what happens when GitHub's auto-SWE is good enough to take on "Write me a GitHub clone, but with additional features X,Y, and Z"? Will they have regret it?
I know this is slightly far-fetched and AI-coders are coming regardless of if GitHub is working on it, but it does seem like these companies are destroying part of their moat (codebase and SW infrastructure). (I also realize their bigger moat is brand and existing user base).
01HNNWZ0MV43FF
If that works, they can always put more money into improving GitHub.
I see this as a technology that "lifts all boats" but like so many technologies, it does lift moneyed boats a lot further.
petercooper
I mean, we have GitLab, Forgejo, Gitea, Sourcehut, and many others all vaguely in that space and they're not eating GitHub's breakfast. If anything, it reinforces just how wide that moat of brand, momentum, and influence/power is, and it might get even wider in an AI world.
ai-christianson
Very cool! I'm working on a similar agent, but FOSS (https://github.com/ai-christianson/RA.Aid) --It'll be really interesting to see how the GitHub agent works.
On first impressions, it looks like they are taking the route of integrating tightly with VSCode, which means they'll be competing with Cline, Cursor, and Windsurf.
IMO it might be good for them to release this on the web, similar to the replit agent. Integration with GitHub directly would be awesome.
spiritplumber
Very neat!
anonzzzies
I downloaded Insiders and installed Github Copilot Chat, getting:
"GitHub Copilot Chat is not compatible with this version of VS Code. Please make sure that you have the latest versions of the extension and VS Code installed."
Piterniel
Same here
isidorn
Sorry about that. Can you switch to copilot chat pre-release extension. Should be a big button "switch to pre-release".
We are tracking this issue here https://github.com/microsoft/vscode/issues/239836
RobinL
Can anyone speak to weather it's worth going back to co-pilot from cursor. On the face of it $10 a month for unlimited messages looks compelling. Is it really unlimited? From these videos it's starting to look pretty similar to cursor...
dedene
I want to know as well. For me, everything I tested so far still can’t beat the autocomplete of Cursor both in speed as intelligence…
> When we introduced GitHub Copilot back in 2021, we had a clear goal: to make developers’ lives easier with an AI pair programmer that helps them write better code. The name reflects our belief that artificial intelligence (AI) isn’t replacing the developer.
Later:
> GitHub Copilot’s new agent mode is capable of iterating on its own code, recognizing errors, and fixing them automatically.
Is this Microsoft/GitHub acknowledging they initially missed the mark, except they aren't really clear in the post that they're abandoning the approach of "AI pair programmer / not replacing the developer"? Seems really strange to re-iterate their "clear goal" and then in the next paragraph go directly against their goal? Are they maybe afraid of a potential backslash if it became more clear that they're now looking to replace (some) developers / the boring parts of development?
I have no opinions either way, but the messaging of the post seems to go into all different directions, which seems strange.