What Is the Trajectory of a Senior Software Engineer in the LLM Era? A Hacker News Debate

A Hacker News discussion titled "Ask HN: What is the current trajectory of a senior engineer?" produced 58 comments and no consensus. The dominant view among commenters is that the senior engineering role is being redefined — from writing code to directing agents, owning quality, and navigating organizational politics — rather than eliminated. A minority view holds that architecture itself is the next target for reinforcement learning and that the role will shrink substantially within a few years.

The question the thread started from

The original poster asked why frontier labs could not simply build datasets of good and bad software architecture and RL on them, the same way they RL on code and math where a right or wrong answer exists. The poster framed the stakes bluntly:

"Given that LLMs were a fancy autocomplete back in 2024, and now people are talking about building 'software factories' with the agents, what in your mind is the stop to this development? Why software architecture would be hard for agents still in a year or two?"

The poster's own conclusion was pessimistic for programmers: "I think (businesswise) the need for actual software programmer skills is 90% out in a few years I think, and product lead etc people will be favored more as they have more people skills. it was a good run."

The strongest counter-argument: architecture may stop mattering at all

The most structurally interesting reply came from akoboldfrying, who argued that the premise of the question is wrong — not because agents cannot learn architecture, but because architecture exists only to work around human cognitive limits.

"Architecture, in the sense of applying/enforcing organising principles like modularity and levels of abstraction, is something we currently do largely to manage complexity -- so that we can fit as much of the system into our limited brains as we can. For a large enough system, this is necessary in order to work on it efficiently. But if you can 'keep the whole thing in your head' without using these 'crutches', then there's actually no need for them. They become overhead that can be done away with."

The commenter compared this to a "Hello world" script, for which nobody plans an architecture. If a model can hold an entire codebase in context, the argument goes, the discipline of architecture becomes optional overhead rather than a skill to be automated.

Evidence from the thread that agents are already close on architecture

One commenter, jwr, reported direct experience using two models named in the thread as Fable 5.1 and Astra on architectural design documents, and said the results were competitive with human work:

"I worked with them on architectural design documents, and they were very good. Some of their suggestions were better than my solutions. They are not at a point where they can fully replace a system architect yet, but they're getting close."

The same commenter identified what remains outside the models' reach: "There is still the open question of what to build and how to create a user interface that is understandable and presents complex information in ways understandable to humans. That's something AI cannot do yet."

A second commenter, zeryx, described attempting to build a video game with agents and reported that the agents could write any individual piece of code, write scripts to catch basic failure modes, and even play the game by spinning up a VM, triggering game state with AutoHotkey, and taking screenshots and GIFs to evaluate what they had built. The failure was not in execution:

"Knowing what to build intuitively to maximise fun, or any other illogical human experience won't be RLable ever."

What the thread says remains hard for agents

Several commenters converged on a set of tasks they consider resistant to automation, and their reasons differ from the usual "AI can't code" argument.

Organizational consensus. accurrent defined the senior role itself as a political one: "IMO senior software engineer is someone who drives consensus in a large organization. I dont think LLMs are going to solve that problem any time soon." The same commenter noted that on the backend, tracing race conditions and reasoning about what fits in a context window still requires human oversight, and that designing tests for seamless migrations does too.

Pushing back on bad ideas. jgilias argued that current models are, if anything, too agreeable to be useful in this role:

"Models have an alignment problem. As in, they are too aligned to push back hard enough on bad ideas. Sometimes you can't really push back because of power dynamics and whatnot, but the job still needs to be done somehow. So humans then navigate it by sometimes just ignoring the bad ideas, sometimes playing along with the narrative, but then doing what actually needs to be done, and sometimes, when there is no other way, digging in heels and killing some bad ideas before they've really taken root. An agent could do this with the SOTA model capability that existed yesterday. But then they'd need to become seriously misaligned."

Accountability. papa-smurf described leaving software development for systems consultancy and named the moat explicitly: "Main moat of the (consultancy) job is that a customer will always need a human person to put the blame on. That's me. I guess similar roles are relatively safe."

Deciding what to build. fergie reduced the question to its essentials: "Ultimately you still need somebody to decide what we are going to tell the machines what to do, and then actually tell the machines what to do, and it doesn't really matter if that is in JavaScript or plain English."

The reframe: senior engineer as technical lead

The most repeated prediction in the thread is that the job title survives but the job description changes. gregwebs put it directly:

"Senior engineer and all SWE positions will become more like a technical lead. If you can direct AI and produce more value per $/time spent then a non software engineer directing AI to develop software then you still can produce a lot of value. This should be the case now for most things that are not a prototype. The job has changed from 'writing code' to 'shipping software'. A technical lead."

The same commenter added a supply-side argument that cuts against simple headcount decline: "But there's also going to be 1000x more software. So there may continue to be opportunities for those that adapt."

bradfa made a parallel point about role convergence, noting that product managers, project managers, and scrum leaders already existed on engineering teams before LLMs, and that those roles were a delegation of tasks rather than a separate species of work:

"Now, all of these roles can be partly offloaded to LLMs, but still having a human in the loop to guide engineering and interface to the business/customer will still be valuable. Knowing what to build overall, what to build next, what's coming later, and how the product relates to the business goals is still very much a human endeavor."

The skills the thread recommends building

RugnirViking recommended a specific technical direction: "I'll suggest more focus on performance, testing, and observability. Human understanding + quality/maintainability is where the money is at. Learn to make those things with agents." The same commenter linked to a separate Hacker News discussion on the idea that "AI usage is a skill," and concluded: "I don't think engineers will truly go extinct, though there may well be less of them."

accurrent pointed to aesthetics and UX intuition as newly valuable on the frontend, because prototyping is now cheap enough to test many concepts, and to robotics as a domain where "there are so many cross functional problems that you dont feel like you are moving faster" even with LLM-assisted coding.

Dissent: the economics are unproven, and the bullishness is self-serving

Not everyone in the thread accepted the premise that agent capability is the binding constraint. sdcfgy argued the constraint is financial and political:

"The economics and the reality are unproven and opaque. Most of testimony we have is heavily invested folk and/or people paying the currently low discounted rates for the technology. The companies providing the technology are not financially healthy, are heavily geopolitically aligned and a risk for many. If it works or not is moot: it looks like a large economic or political failure on the horizon. Ergo wait and see before thinking about the future."

amelius framed adoption itself as a trap: "Every vibe coder here that is bullish on AI needs to realize that we're training the AI. Helping it to replace us. We're basically turkeys voting for Christmas. We're cooked."

fergie offered a competing explanation for any downward pressure on senior salaries: "There does seem to be some pressure on the top end of SWE salaries. Its possible that this is just wishful thinking, but if true, I suspect that it has more to do with an increase in job applicants rather than a decrease in actual work. AI seems to have driven up the amount of applications received for every vacant position."

zerop made a narrower claim about where adoption will actually land: "LLMs have been able to make good success in coding, because someone is reviewing the work it produces (code). Ofcourse quality of code has become better over time. IMO other than Human-in-the-loop scenarios, AI will not see huge adoption in businesses."

The demand-side argument: most people do not want to talk to computers

dimbletimbers argued that the thread's participants systematically overestimate the competition:

"I think devs and people on HN way overestimate how much the average person knows about or cares about or desires to interact with a computer. Whatever the job becomes, if you like talking to the computer and know how to get it to do anything besides show an email or youtube, you're ahead of the curve and can likely stay there."

testerius extended the point to a general labor argument, placing programmers alongside mathematicians, doctors, biologists, sales representatives, marketers, teachers, and artists: "Of course you can get rid off programmers or mathematicians but then you need to understand and handle stuff on your own. You still need somebody who can check 'what's going on'."

frangonf made the minimal version of the same case: "if software (and rest of disciplines possible soon to be hit as hard) can still be 'wrong', if you can still mess it up even with machine assistance, people doing software still have a place."

Where the thread lands

No commenter claimed to know the timeline. The thread's actual output is a list of candidate moats — organizational consensus, accountability, deciding what to build, aesthetic and UX judgment, performance and observability work, and the willingness to overrule a bad idea — alongside two genuinely incompatible predictions. One is that architecture is a learnable skill that agents will absorb, after which the discipline becomes unnecessary overhead. The other is that architecture is a workaround for human memory limits that agents never needed in the first place, so the question of whether they can learn it is beside the point.

Several commenters responded to the existential framing with dark humor rather than analysis. Zealotux wrote: "I've started looking into creating a pest control business, or maybe renovating furnitures I find on the street. Let me know if you have other ideas for senior engineers." sitzkrieg, self-described as "almost officially graybeard territory," said: "thinking about starting an apiary. getting stung by 100s of bees still sounds like more fun than being forced to have to interact with LLM code from any source." sharpshadow offered the most compressed version of the thread's anxiety: "Coding by hand will be prohibited. Too big of a risk factor. You may be the shaman communicating with the almighty."