How Software Actually Gets Built — and What AI Just Changed

Films sell the lone genius in a hoodie; the industry runs relay teams. The Software Development Life Cycle — SDLC — is the loop every product you use runs on: plan what to build and why, design how it works, build it, test it, ship it, watch it in production, and go again. The banking app you tapped with this morning, Uber Eats, WhatsApp — each release is the work of dozens to hundreds of specialists, and the loop turns every week or two. Hold this loop in your head for the whole session. The next fifty years of history are the story of one industry learning to turn this wheel faster and with quicker feedback — and the AI section at the end is simply the newest, fastest turn of the same wheel.

In 1970, Winston Royce published the paper that gave us the waterfall diagram — requirements, design, code, test, ship, each phase finished before the next begins. The twist: in the same paper he called that exact approach 'risky and invites failure' and argued for doing phases twice with feedback loops. The industry adopted his diagram and ignored his warning — the most misread paper in software history. The bill arrived in the 1994 Standish CHAOS report: 16.2% of projects succeeded, 31.1% were cancelled outright, and the average cost overrun was 189%. The FBI's Virtual Case File is the poster child — scrapped in April 2005 after roughly US$170 million, because 700,000 lines of code met their users years too late. Waterfall's fatal flaw: feedback arrives after the money is gone.

February 11 to 13, 2001, the Lodge at Snowbird, Utah. Seventeen software rebels — Kent Beck, Martin Fowler, Jeff Sutherland among them — skied, argued, and wrote four values and twelve principles that fit on a single page: agilemanifesto.org. They rejected the label 'lightweight' because nobody wants to be a lightweight, and, per Jon Kern, 'we left our egos at the door.' The roots ran older — a 1986 Harvard Business Review paper about rugby-style product teams, the first Scrum at Easel in 1993. The payoff shows in Standish's later data: agile projects succeed around 42% versus roughly 13% for waterfall. Best redemption story: the FBI's Sentinel system floundered until 2010, then cut the team to about 45, adopted two-week sprints, and shipped bureau-wide on July 1, 2012 — under budget. Same bureau that lost $170 million; new loop.

June 2009, Velocity conference: John Allspaw and Paul Hammond of Flickr present '10+ Deploys per Day' — when the industry norm was quarterly releases. Patrick Debois watched remotely and ran the first DevOpsDays in Ghent that October; the word 'DevOps' exists because 'Agile System Administration' was too long for a Twitter hashtag. By Velocity 2011, Amazon revealed it deployed to production every 11.6 seconds on average. Why the obsession with automated pipelines? August 1, 2012: Knight Capital manually deployed to seven of eight servers, dead code woke up behind a reused flag, and $440 million evaporated in 45 minutes — the firm was sold by December. No automation, no kill switch. Yet DORA 2025 finds only 16.2% of organisations deploy on demand and 43.5% still take over a week. Elite practice is rare — that gap is your opportunity.

Let's slow the Knight Capital tape down, because the mechanics teach more than the number. Engineers hand-copied new routing code to eight servers over several days — and reached only seven. The rollout reused an old feature flag, 'Power Peg', still wired to a retired test routine built to buy high and sell low on purpose. At the market open on 1 August 2012, orders hitting the eighth server woke that dead code, firing millions of orders exactly backwards — and with no kill switch, the confusion ran 45 minutes. Map each failure to its fix and you get the DevOps checklist: scripted deploys that treat every machine identically, flag hygiene that deletes retired code, instant rollback, alerts tied to business metrics. Keep the list — 2025's Replit database deletion is the same failure class wearing AI clothes.

Every logo on this grid started as somebody's specific pain. Start with the one that is local: Sydney, 2002, two 22-year-olds fund Jira on about ten thousand Australian dollars of credit-card debt and name it after Gojira — Godzilla — as a jab at Bugzilla; Atlassian booked US$5.22 billion revenue in FY2025, built on a self-service model that famously grew for years with no traditional sales force. The board you will run your sprints on was built here in Australia. April 2005: the Linux kernel loses its free BitKeeper licence, so Torvalds starts Git on April 3 and it hosts itself by April 7 — about ten days to usable. Slack is the salvaged chat tool of a dead game; VS Code's ancestor Monaco had about 3,000 monthly users before the 2015 pivot — 75.9% of developers use VS Code today. And Octoverse 2025's headline: a new developer joins GitHub every second, and AI-assisted work pushed TypeScript to the number-one language on the platform.

Follow one feature through a real team. A product manager writes the ticket with acceptance criteria; a designer draws the Figma flow; you branch and code; a teammate reviews your pull request line by line; CI and QA prove it works; DevOps ships it behind a feature flag and dashboards watch it live. LinearB's 2025 benchmarks — 6.1 million PRs across 3,000 organisations — put elite review pickup under 7 hours and full PR cycle under 26, with small PRs the number-one velocity driver. Now the myth-buster: Microsoft's Time Warp study clocked the average developer week at roughly 12% meetings, 11% coding, 9% debugging. Atlassian's 2025 survey of 3,500 developers found half lose ten-plus hours a week to friction like hunting for information. Remember this when AI enters the story: typing was never the bottleneck.

Before tonight's field study, let's read one merged PR together so the timeline feels familiar. Every PR on GitHub has the same skeleton: a description at the top, then interleaved events — commits, review comments, CI check runs — ending in a merge. Walk this one: opened Tuesday morning with 'Fixes #212345', so the work is traceable to a reason; fourteen checks pass across three operating systems before any human looks; a reviewer pins a question to line 87 — comments only, never editing the author's code; the author pushes a fix; approved and merged Wednesday. That rhythm is exactly LinearB's elite band — pickup under 7 hours, cycle under 26. And a PR with no comments merged in three minutes tells you something too: solo maintainer or rubber stamp. Reading these threads is free practice in the reviewer's eye this era pays for.

Autocomplete is older than you — IntelliSense shipped in 1996. Kite built AI completion from 2014, reached 500,000 developers, and shut down on 21 November 2022; its founder said they were 'ten years too early' — nine days before ChatGPT launched. GitHub Copilot went GA in June 2022 at $10 a month, already writing about 40% of new code in enabled files. ChatGPT hit 100 million users in two months, and Stack Overflow's monthly questions have since collapsed by roughly three-quarters. February 2025: Karpathy's 'vibe coding' tweet became Collins' Word of the Year within nine months. Claude Code went GA in May 2025 and passed $1 billion annualized revenue in about six months; Cursor 2.0 now runs up to eight parallel agents. Thirty years from autocomplete to agent fleets — with the whole inflection packed into the last five.

Mid-2026, stage by stage: the loop from our second slide is intact, but every stage grew an agent. Planning: over 70% of product managers draft PRDs with AI daily. Design: Figma's 2026 survey has 72% of designers using generative AI. Coding: about 27% of production code industry-wide is now AI-authored, measured across 4.2 million developers — Microsoft says around 30%, Anthropic 70–90%. Review: 44% of teams run AI code review, and CodeRabbit, after 13 million PRs, finds AI-coauthored PRs carry about 1.7 times more issues. Operations: PagerDuty's SRE agent can literally join the on-call rota. One cautionary tale to keep: in July 2025 a Replit agent deleted a live production database during a code freeze. The human job moved from producing to specifying and verifying — with guardrails, always.

Both columns are true; the variable is context. February 2023, controlled trial: 95 developers building a greenfield HTTP server were 55.8% faster with Copilot. July 2025, METR's randomised trial: 16 experienced open-source developers, 246 tasks in codebases they knew deeply — 19% slower with AI, after predicting 24% faster, and they still believed afterwards they had been 20% faster. A February 2026 follow-up with 57 developers landed near zero — neither miracle nor catastrophe. You cannot feel your own productivity; measure it. Quality: Veracode found 45% of AI-generated code fails OWASP security tests, flat for two years; GitClear logged 2024 as the first year copy-pasted code exceeded refactored code. DORA 2025's verdict resolves the paradox: AI is an amplifier — teams with tests, CI and small batches compound the gains; weak teams ship accelerated chaos.

Straight numbers, no doom — and be honest that the best-measured numbers here are American, because that is where the payroll research exists. Stanford's payroll study found employment of 22-to-25-year-old US developers fell about 20% from the late-2022 peak while older cohorts stayed flat. We have no equivalent Australian payroll study, so treat that as an overseas signal and check the local picture yourself: Jobs and Skills Australia publishes occupation shortage analysis for software and applications programmers, and the ABS Labour Force release gives the underlying employment trend. That entry-level signal is also not the whole story: postings that ask for AI skills advertise a pay premium. The quotable premium figures circulating are overseas ones, so we are not going to put a number on the Australian premium — read the local ads and see for yourself. Whatever it turns out to be, it is paid on top of fundamentals, never instead of them. The new titles are real — AI Engineer, forward-deployed engineer (up 800%+), agent ops — and each decodes to an SDLC role wearing new clothes, not 'prompt typist'. One habit worth building tonight: take any two current graduate or associate-engineer ads — say one from a bank's technology division, one from a consultancy — and strip each into fundamentals, AI-workflow, evidence. Hiring managers open your GitHub before your resume. The seats moved. Follow them.

Give this ten minutes; there are no wrong answers. First: would you merge a pull request your friend generated entirely with AI — and what would you check before merging? Listen for verification instincts and connect them to CodeRabbit's finding that AI-coauthored PRs carry about 1.7 times more issues. Second: which era would you rather have started in — 1995, 2010, or now? Expect a real split: some crave the perceived purity of pre-AI coding, others the leverage of today. Third: what would you still learn deeply even though AI can generate it? Steer the close towards DORA's amplifier idea — fundamentals decide whether AI multiplies you or your mistakes. Non-CS students are welcome voices here: employers increasingly want domain knowledge plus AI skills, so a mechanical engineer who can verify AI output is exactly the new profile.

Run this as a show of hands or shout-outs — no marks, no gate, purely a pulse check. One: waterfall's fatal flaw in one word — feedback, arriving years too late. Two: what did Flickr do ten-plus times a day in 2009 that shocked the industry — deploy to production. Three: the METR trial — were experienced developers faster or slower with AI? Slower, by 19%, while believing they were 20% faster; anyone who confidently answered 'faster' has just demonstrated the perception gap live. Four: what fraction of a developer's week is actually spent coding — about 11%, per Microsoft's own Time Warp study. Anyone who got all four right can already explain this industry better than most of LinkedIn. If any question felt fuzzy, the recap on the next slide is your one-minute revision.

Four things to carry home. One: software is a loop — plan, design, build, test, ship, watch — run by a team, and the loop matters as much as the code. Two: every era fixed the previous era's feedback pain: agile shrank years to two weeks, DevOps shrank weeks to minutes, and the disasters — a $170 million scrapped FBI system, $440 million lost in 45 minutes at Knight Capital — are exactly why those practices exist. Three: AI, from Copilot in 2021 to agent teams in 2026, moved the developer's centre of gravity from typing to specifying and verifying — and typing was only ever 11% of the week anyway. Four: the best evidence says AI is an amplifier; your fundamentals and process decide whether it multiplies your output or your chaos. Verification is the scarce, hireable skill.

Close the loop by sending the room somewhere good tonight — pitch these as the primary sources behind today's stories, not as homework. Karpathy's 'Software Is Changing (Again)' keynote is the big-picture map behind our 'Then AI arrived' arc — the best forty minutes a curious student can spend this week. Fireship's hundred-second CI/CD video makes the yearly-releases-to-11.6-seconds jump click visually. Doug Seven's 'Knightmare' is the minute-by-minute Knight Capital retelling behind our anatomy slide. Dare them to read Royce's 1970 paper — page two alone proves the warning was real — and the Agile Manifesto takes two minutes at the source. DORA's site hosts the amplifier evidence. Tell everyone the full clickable list, with a why for each pick, lives on this session's LMS page — no scribbling URLs from the screen.