I am pretty familiar with a 500k LOC codebase. If for every feature request/bug the agent has to go through a lot of it, spend a gazillion thinking tokens for understanding what it needs to do, plan, and then execute (assuming it gets it right) given the current cost of tokens I argue I am often more cost effective.
In fact, I believe that the most cost effective way is a collab of human+agent. Ie giving the agent direction as it goes along with the plan I can cut the thinking while keeping the speed. Basically helping the agent going from a breadth first search into a guided depth first one which is much more token efficient.
Additionally, humans have long term memory and knowledge of the context around your codebase. Agents do not, and while you can fit a lot in 1M context window, once you fill that the quality goes down considerably.
Not to mention the fact that even the most well documented codebase will have documentation blindspots about real-world concerns or limitations that LLMs cant know about. Cursor yesterday tried to remove a document format from the codebase because it was convinced that it was non-existant, turns out that not only does it exist and is vitally important for our shipping process, but also the API it comes from does not document its existence at all.
This is why you can't be replaced today. I'm not sure you can rely on that remaining true for very long. And this goes for the vast majority of us.
To be clear, I'm also not saying LLMs will definitely displace a lot of us very soon. I'm just saying I wouldn't be surprised by either outcome and I don't know how anyone claims to know one way or another given the past year or so of progress.
Im curious if that point comes before it automates away the entire mid-upper management caste.
In a hypothetical world where LLMs have enough context window and "understanding" to have no need for an experienced user to give inputs I would assume its also going to have enough information to make most business decisions and provide well formatted info to the C-Suite.
I think you are assuming cost per task will become cheaper and that there is unlimited energy supply.
While tokens costs are going down, the number of token burned is going up and up. Case in point Sam Altman is complaining about their top token users burning through 100B tokens per month [1]. So you have token prices going down but token usage going up 10x per year (if you extrapolate linearly from what Sam was ranting about). This is happening because people trust more and more LLMs and give them more autonomy and more complex tasks (IMHO).
So if you really need a true unsupervised agent that replaces SWEs you need how probably much more than that. Say 20x that number (2T tokens/month) for each SWE. I'm gonna focus on the energy part as this is more tangible. Trying with some realistic numbers:
- To replace 1M SWEs for a year you need 2T tokens/month * 12 months * 1M SWEs ( = 2.410^19 tokens)
- Assuming 0.5J per token you get 1.210^19J [2] (I took the number for an llama3 8B model, probably is much more for SOTA models IMHO).
- A year has 31M seconds
- Over a year that is 380 GW of constant power that is needed only for replacing 1M SWEs and that is around 80% of all the current US energy consumption (450GW). And apparently there are 47ish Million SWEs globally as of 2025 [3]
I don't think there is enough power capacity to deliver all of this without pivoting all of society into building data centers and power plants.
So unless there is some breakthrough in efficiency/intelligence (ie you need way fewer tokens for what you have to do) your job is gonna be safish at least.
Of course I pulled that 20x out of my ass, but I believe it is somewhat realistic for a truly autonomous agent(s) that replace SWEs.
> - Over a year that is 380 GW of constant power that is needed only for replacing 1M SWEs and that is around 80% of all the current US energy consumption (450GW). And apparently there are 47ish Million SWEs globally as of 2025 [3]
I think the economics here work out as "OK, so we've bought 80% the electricity in the US and used this to sell software to the 96% of humans not living in the US; this is profitable for the businesses, so nobody with money cares about the Americans who now literally can't afford to keep refrigerators running because we outbid them".
Someone is not let go with the announcement "you are replaced with AI". I know many teams that have downsized, or are not hiring after someone left. There is a reason why leverage of employees has drastically gone down. I myself am struggling in this aspect.
I think the missing piece here is nuance. Of course there are certain tasks that software engineers do that will be replaced. But will AI replace _everything_ a software engineer does?
The most difficult bit about software engineering is to keep a mental model of _everything_ a product does with varying levels of granularity. The way I see LLMs fail at my company the most is that they are very good at the big picture, and very good at the very small picture, but have difficulty moving between those two levels. And especially when changes have occurred or accumulated over time. Most of all production systems have an extremely long tail of gotchas which are only managed by people who have been around for long enough to have some kind of deep storage access in their heads to those little tidbits of information.
And I think current LLMs might be fundamentally incapable of replacing that.
Maybe tomorrow, but they're just still not there quite yet. For non-software developers, there's a "vibe coding wall" that gets hit and the project needs a software developer to unwind. Maybe Mythos or codex-5.6 will be able to do it, but Opus 4.8 and codex-5.5 still get stuck without proper guidance.
Having long time perspective, seeing the wider context, talking to people, having vision and taste. Curating and sanity-checking the results. Security and trust.
I've yet to hear an argument that argues that software engineers can be replaced by AI that doesnt boil down to slop apologism, inability to detect slop or simple gaslighting.
These are things I've come to expect from bots, clueless journalists, clueless juniors, clueless expert beginners and clueless members of the professional managerial class but almost never from experienced software engineers.
To be fair, seasoned software engineers always seem to get shouted down online by the former group which is louder and more numerous so you could argue that we "lost" the argument.
Meanwhile big tech's vibe coded monstrosities are increasingly exploding all around us in ever more humiliating ways while the humans who had this tech rammed down their throats get thrown under the bus.
This undeserved halo effect over AI is maintained in order to keep the needle from pricking the ginormous stock market bubble that hinges upon the religious belief in the lie AI Will Replace Us All Soon.
The argument is "Software Engineer" sounds like "Programmer" to me and "Programming" is just typing lines of code, AI can do all that typing quicker than a human so there we go.
Currently leading an Integration that for the most part needs no new code written and the CEO is breathing down my neck telling me to cut my 4 week estimate down to 1 because "can't i just use AI like the other firms do?".
There's a morbid part of me that wants to give him what he wants and let claude make critical process decisions on internal processes that are very domain specific and have no online documentation, but alas I would rather not have the project go down in flames so I smile and nod.
Much less emasculating than accepting that all that weird tech mumbo jumbo that those overpaid senior engineers babble about actually has a deeper meaning and is not just there to artificially slow down the all important company growth!
> I've yet to hear an argument that argues that software engineers can be replaced by AI that doesnt boil down to slop apologism, inability to detect slop or simple gaslighting.
That, or extreme extrapolation from events that form a vanishingly tiny part of the job of a software engineer. "Last week my AI solved this amazing software problem that I had struggled with" very quickly becomes "the AI is better at software than I am". Any pushback suggesting that the fact that something (or someone) did one tiny part of your job better than you one time does not mean you should be replaced, is quickly met with "yeah, but that's today, imagine how amazing the models will be in n years".
You can't win a debate with this much moving of goalposts.
In fact, I believe that the most cost effective way is a collab of human+agent. Ie giving the agent direction as it goes along with the plan I can cut the thinking while keeping the speed. Basically helping the agent going from a breadth first search into a guided depth first one which is much more token efficient.
Additionally, humans have long term memory and knowledge of the context around your codebase. Agents do not, and while you can fit a lot in 1M context window, once you fill that the quality goes down considerably.