The New Software Engineer
Coding, AI, Agents, and What Comes Next.
Hi, I’m Bryl.
AI systems by day, apps and hackathons by night. I started exactly where you are.
- Principal AI EngineerStandard Chartered
- Ex-GenAI Center of ExcellenceCambridge
- Software Engineering LeadPocketDevs
- FounderAppBuildersPH
What we will cover.
One question for the whole hour: if the AI writes the code, what is the engineer for?
Show of hands: who used AI to write code this week?
The job used to be typing code.
Typing code is now the cheapest part of the job. Deciding what to build, and checking it, is the job.
Engineer as typist
- Memorize the language
- Search, copy, adapt
- Hours of repetitive setup code
Engineer as director
- Explain the problem clearly
- Say what “done” looks like
- Check the result
From my workAt PocketDevs I led 60+ app builds with a team of ten. Most of those hours went into typing repetitive code.
Three eras in ten years.
What a new engineer is expected to do on day one keeps changing.
The computer only checks your spelling.
It finishes your line for you.
You give a goal. It reads, edits, tests, and reports back.
More code than ever. More coders than ever.
AI did not shrink the field. It made it bigger.
So the question is not whether there are jobs. It is what the job is.
GitHub Octoverse 2025
Almost everyone uses AI to code now.
Two large surveys of working developers, same answer.
At work, using AI is expected. Using it well is the skill.
Stack Overflow 2026 (30,000+ devs) · JetBrains 2026 (15,000+ devs)
How much code do agents write?
Developers were asked how last month’s code was written. Self-reported, so treat it as a rough picture.
The groups overlap, so the numbers add past 100%. Karpathy says AI writes about 80% of his own code.
Has an AI ever given you code that looked right and was wrong?
What an AI model actually does.
It guesses the next word piece (a “token”), adds it, and guesses again. That is the whole trick.
Three ways to program a computer.
Karpathy’s idea. All three are still in use today.
You spell out every step. The computer follows exactly.
You give it data. It learns the rules itself.
You describe what you want in English. The AI works out the steps.
Smart and unreliable at the same time.
Karpathy calls it “jagged intelligence”: amazing at some tasks, surprisingly bad at nearby ones. Knowing which is which is your first real skill.
Things it has seen a million times
- Repetitive setup code
- Converting code from one language to another
- Explaining code you do not understand
Things it cannot see or check
- Your project, unless you show it
- Knowing when it is wrong
- Deciding what is worth building
From my workTunai, an AI fact-checker I built, exists because of this. The model argues well but checks nothing unless you make it.
The AI only knows what you show it.
Give it the goal, the right files, the real errors, and current docs. Most “bad AI” moments are really missing context.
From my workAt Cambridge I built a WhatsApp lesson planner for South African teachers. The trick was not the model. It was showing it the right curriculum pages. Lesson prep time fell 70%.
Everyone uses it. Few trust it blindly.
The gap between using AI and trusting it is where your job lives.
Only 10% use AI answers without checking. Stack Overflow 2026.
From my workAt Standard Chartered nothing I build goes live without automated checks and a human sign-off. A bank cannot run on 7%.
Feeling faster is not the same as being faster.
The only controlled experiment so far found the opposite of what developers felt.
METR, 2025. A 2026 follow-up was inconclusive. Nobody has a clean number yet. Be careful with anyone who claims one.
If you could hand one task to an AI agent this week, what would it be?
From chatbot to agent.
A chatbot answers. An agent does: it thinks, acts, checks the result, and repeats until the job is done.
From my workeGov Agent, our AI agent for Philippine government services, is this loop pointed at a government form: find it, fill it, check it.
Agents use the same tools you do.
Nothing magic. It runs the same commands you type. Better tools and tests make a better agent.
From my workAt Cambridge I connected an AI model to Zendesk so support staff could ask for shipment status in plain English. The model was the easy part. Connecting it was the work.
How much should the AI do? You decide.
Karpathy’s “autonomy slider”: from suggesting one line to building a whole feature. Slide it up as the AI earns your trust. Iron Man suit, not Iron Man robot.
Vibe coding vs. agentic engineering.
Karpathy coined “vibe coding” in 2025: let the AI write it, do not read it. In 2026 he said professionals need something stricter.
“Fully give in to the vibes”
- Prompt, accept, move on
- Great for quick prototypes
- No quality bar
“Coordinating fallible agents”
- Write the plan, review every change, run checks
- Keep it correct, secure, maintainable
- A human stays responsible
Vibe coding raises the floor. Agentic engineering is about extrapolating the ceiling.Andrej Karpathy · Sequoia AI Ascent · April 2026
The discipline has four parts.
Nothing new. These are the habits of a good senior engineer, applied to a fast worker that is sometimes confidently wrong.
From my workThis is my day job at a bank: written plans, every change reviewed, automatic checks, tight permissions.
Agents fail in boring ways.
I have seen each of these on a real project. The fix is good habits, not better prompts.
Confident and slightly off
- Calls a function that does not exist
- Makes the test pass by deleting the test
- Fixes one bug, rewrites six other files
Small loops, hard checks
- Tests it must pass but cannot edit
- Small tasks, small changes
- Read every change before merge
A working loop with a coding agent.
Andrew Ng’s checklist, September 2026. Short supervised steps beat long unsupervised runs.
Not the year of agents. The decade.
Karpathy’s take: agents are real but far from reliable. Going from 90% to 99.99% reliable takes years.
The problems are tractable, they’re surmountable, but they’re still difficult. If I just average it out, it just feels like a decade to me.Andrej Karpathy · Dwarkesh Podcast · October 2025
Good news for a 1st year: the field is still being built. You arrive in time.
What can you do in four years that an AI cannot?
The four skills that matter.
Andrew Ng’s AI Engineering Skills Map, August 2026. Writing prompts is a small part of one of them.
Test them properly. Learn from the errors.
Know the tradeoffs so you can steer the AI.
What to show it, how much to let it do.
The AI follows a clear plan. Someone has to write it.
From my workI ran a company-wide AI bootcamp at Cambridge. The gap was never prompting. It was fundamentals and knowing what to build.
Fundamentals matter more, not less.
You cannot check what you do not understand.
Developers who deeply understand how software works vastly outperform those who vibe code without understanding.Andrew Ng · Skills Map, Part 2 · August 2026
You do not have to build every layer yourself. You have to be able to judge what the AI built.
From my workMyCambridge serves 500,000+ users. When it slowed down, the fix was a database index and a cache, not a prompt.
If the AI writes the code, what do you do?
You move one seat over. Less typing, more deciding.
Turn a vague request into a clear goal.
If you cannot describe it, the AI cannot reach it.
Read the change. Catch what the tests missed.
“The AI wrote it” is not an excuse in production.
The new job description.
Companies still hire engineers. They pay for what the AI cannot do.
Writing code is the entry ticket. Judgment is the job.
Who gets hired in 2026.
Juniors are not finished. Juniors who work like it is 2022 are.
I have often hired a fresh grad who knows AI over a veteran developer who is still coding like it’s 2022.Andrew Ng · interview with Salesforce · 2026
From my workI built CodeCred for exactly this: proof of skills you can verify, instead of a line on a résumé.
What year are you in?
A plan for wherever you are.
Build something. Find what you do not know. Learn it. Build again.
Struggle on purpose. Use AI to explain, not to answer.
Pair with a model. Read everything it writes.
Join hackathons. Build whole projects. Write plans an AI can follow.
Depth is what makes your judgment worth paying for.
From my workCapstone Generator was a year-3 kind of project: a tool for IT students, shipped, and still used.
Hacks for learning with AI.
Use the AI as a tutor, not a vending machine.
Hacks for building with AI.
Habits that keep the agent useful and you in control.
Ship things. Hackathons compress years.
Every role on my intro slide traces back to something I built and showed someone.
In 48 hours you learn what a semester cannot teach: scoping, shipping, and explaining your work to strangers.
Tarsi#1 Finance app · offline, on-device AI
Kabi#1 Paid Fitness app
Mayi#1 Paid Travel appThree loops to run every week.
Andrew Ng’s three feedback loops: fast, medium, slow. Run all three on your own projects.
Plan, run, review, save.
Use it yourself. Fix the plan, not just the code.
Show it to strangers.
What comes next.
Nobody knows the details. The direction is clear enough to plan for.
From my workeGov Agent already works this way: one agent reads the form, one fills it, one checks it. I review the result.
Learn the fundamentals. Use the tools. Ship.
Learn the fundamentals. Use the tools. Ship.
The tools will keep changing. The engineer who understands the system and ships anyway will not go out of style.
bryllim.com/presentation
Sources.
- R-01Stack Overflow, 2026 Developer Survey, AI chapter. survey.stackoverflow.co/2026/ai. Released October 2026.
- R-02JetBrains, State of Developer Ecosystem 2026. “How much code do developers really let agents write?” blog.jetbrains.com/research, August 2026.
- R-03METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” July 2025, and the February 2026 follow-up on changing the experiment design.
- R-04GitHub, Octoverse 2025, github.blog, October 28, 2025.
- R-05Andrej Karpathy, “Software Is Changing (Again),” YC AI Startup School, June 2025; Dwarkesh Podcast, October 2025; Sequoia AI Ascent fireside chat and blog summary, April 2026.
- R-06Andrew Ng, “AI Engineering Skills Map,” deeplearning.ai, August 14, 2026; Part 2 on software fundamentals, August 29; coding-agents deep dive, September 4, 2026; interview with Salesforce, 2026.