every year AIE has its own meta.
last year: evals, coding agents, and a third thing i genuinely already forgot (that's the point)
this year: software factories. multiplayer ai. local models.
next year: something else. we'll all chase it.
biggest unlock from this week: there's already SO much shipped - oss market - that the play might not be writing everything yourself anymore. it's combining what exists, with your taste as the glue.
as an engineer this is the most painful pill to swallow. "i'll just write it myself" is the default cracked-dev bias. but i think it's the way.
my last 12 months: built my own orchestration framework. my own harness. contributed to openclaw. all of it leveled me up - and exposed the pattern: i was building for imaginary users with imaginary needs, instead of a problem i hit every single day. now i know exactly what to build.
next 12 months: DEVBOXES.
the pitch is stupid simple:
your agents should not run on your laptop.
they live in an isolated remote box with full permissions - actual yolo mode, no approving sketchy commands every 2 minutes. assign long tasks, close the laptop, get pinged when they need you or they're done.
a remote vps with coding agents pre-installed. that's it. explainable in one breath.
and that matters. half the expo floor was products so packed with features you couldn't find the golden path. i get it - stack features so the big labs can't eat you in one bite. i'm going the other way: one simple, robust thing.
simple ≠ easy. simple is hard mode. complexity is what happens by default when nobody makes decisions. simplicity IS the decisions - what to cut, what to say no to, which single path to perfect. complicated-but-impressive is cheap. powerful-but-obvious is real engineering.
btw a devbox is the minimum viable software factory. add multiplayer local models on top and we're fully trend-compliant for 2026 :)
new term unlocked: REMOTE CODING.
ps: watch
@theo's closing keynote - the whole "unlearn your old principles" thing is exactly this.
