In early June, AI Engineer Melbourne brought together close to ninety speakers, across two days and four tracks, on just about every aspect of building with AI. This is a report on what they had to say — every claim linked to the talk it came from, so you can check for yourself.
With four tracks running at once, nobody at the conference — me included — saw more than a fraction of the talks. This year, for the first time, we kept a full record: every session recorded, every slide captured, and nineteen speaker interviews after the event. That makes it possible to look at what the conference had to say as a whole, rather than the quarter of it any one of us sat through.
A few things stood out. Speakers in different rooms kept making the same arguments — sometimes nearly word for word — without knowing it. They also disagreed, often in useful ways. This report gathers all of that into six themes, with the evidence for each, the open disagreements, and questions worth discussing with your team. Every quote and claim links back to the talk it came from.
How it was made — plainlyThere's more material here than one person can honestly work through — around fourteen hours of transcripts, thousands of slides, nineteen interviews — so an LLM did the reading and the first drafts, and I've done the framing, the judgement calls and the editing. Worth knowing, and it's also why everything links to its source: you can always see what was actually said.
What's in it — and what to do with itThe clearest thread across the two days: models are becoming components. George Cameron opened the conference with market data showing the gap between frontier models narrowing and prices falling fast, and speakers across the program — most of whom saw none of the others' talks — kept making the practical version of his point: match the model to the task, route between models, and know when a small model or a hundred lines of Python will do. This has happened before, with servers, browsers and frameworks. Once something becomes a commodity, the interesting engineering moves to what surrounds it. 01 · The harness beats the model →
That shift raises problems the speakers kept returning to. Cost: token spend is now an engineering concern, not just a finance one — Sarah Sachs argued compute strategy is product strategy; Geoff Huntley wants AI run like a factory rather than a vending machine. 05 → Memory: agents forget everything between sessions, and a bigger context window doesn't fix that — Igor Costa's keynote and several of the agent talks hit the same wall, and this one's genuinely unsolved. 02 → Testing: you can't ship non-deterministic systems on vibes, so evals and observability are taking the place tests and ops took a decade ago — REA took a five per cent hallucination rate in front of real users down to zero reported, by building the eval system first. 03 → And the newest idea of the conference: if code can be regenerated on demand, what's worth preserving is the specification — “we've been backing up the wrong files.” 04 →
From the Leadership track, the theme I suspect matters most: the limiting factor isn't the technology, it's organisations — governance, trust, culture, and how it feels to be twenty years into a career and worried about being junior again. 06 →
The speakers didn't agree on everything. Are harnesses where the value is, or already overbuilt? Do coding agents change everything, or not much? I've kept those disagreements visible in each theme, along with questions worth raising with your own team — they're the same questions most teams are wrestling with right now.
All six themes are written up in full below, evidence and all. We'll do this again for AI Engineer Sydney in December.
The base model is a commodity; the moat is routing, context, cascades — and knowing when a smaller model (or none) does the job.
Read theme 01 →Agents forget everything, and a bigger window doesn't fix it. Where memory should live — files, platforms, or the model itself — split the speakers three ways. Honestly unsolved.
Read theme 02 →You can't ship non-deterministic systems on vibes. Deterministic checks first, judges on a leash, traces as the shared language.
Read theme 03 →“We've been backing up the wrong files.” When code is regenerable, what do we keep — the spec, the loop, or the why? One room, one afternoon, three answers.
Read theme 04 →Token spend is engineering now. Prices per capability collapse; total spend rises anyway — both true at once, and the same cautionary tale surfaced in two rooms.
Read theme 05 →The limiting factor is organisations — governance shape, boardroom miscalibration, and engineers afraid of becoming junior again.
Read theme 06 →