▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=Ru99FGJ_yuE
• Original title :: The state of the tech industry in 2026 - LDX3 New York

Gergely Orosz's LDX3 New York keynote (15 Sep 2026), based on visits to OpenAI, Anthropic, Cursor, Ramp, Linear and Uber. In twelve months AI went from assistant to near-100% code author: parallel agents, fading IDEs, in-house harnesses and agentic software factories are the norm. Teams, planning and testing still matter; code review, reliability, CPU capacity, focus and engineering leadership are breaking. His advice to leaders: get and stay hands-on, build team-level AI infra, and expect less people management for now.

What changed

• Nobody writes code by hand :: Since Opus 4.5 / GPT 5.2 (January), most startups report close to 100% AI-generated (human-prompted) code.
• Parallel agents :: Boris Cherny runs 5 local + 5–10 web Claude sessions; Linear's Dima and Cockroach Labs' Peter Mattis routinely run 5–10 worktrees/agents.
• IDEs fading :: Windsurf/Antigravity was the last big VS Code fork. OpenAI almost forked VS Code for Codex but didn't, and was glad. JetBrains pivots to Air; Cursor pivoted to an agentic UI in April and calls the IDE a legacy enterprise product. Kent Beck: we still need context for human decisions, but the context changed.
• Everyone builds a harness :: Ramp, Stripe, Uber, Block, Shopify, Google, Meta, Amazon, DoorDash, Monzo and others build their own agent platforms, often on Codex / Claude Code / OpenCode, to plug in internal data.
• Work starts in Slack :: Tag Codex/Claude/Linear bot to implement something.
• Agentic software factory is real :: OpenAI's agents react to signals and raise PRs, including a "perf factory" that watches production; humans review high-blast-radius PRs.
• Migrations take months, not years :: Bun's Zig→Rust port by one engineer; OpenAI's API Python→Rust 90% done in 4–5 months; Airbnb Enzyme→RTL in 6 weeks; Asana in 2 weeks; Uber JUnit 4→5 (millions of LOC) in 4 months.
• Agent-created work overtakes humans :: On Linear, agent/MCP-created issues now outnumber human ones. GitHub agent-only PRs grew ~10x Jan→Aug (7.7M in January). Agent-only-reviewed PRs spiked in Jul–Aug.
• Skills mainstream :: Factory AI sees 80%+ of users using skills, up from ~30% in February.
• Cost :: Spring panic (Uber blew its annual AI budget). By September large companies cut per-token cost ~50% via open models and smarter routing.
• Specializations disappearing :: OpenAI stopped hiring dedicated iOS/Android engineers; frontend engineers do backend.
• Projects of 1–2 engineers :: Anthropic: more engineers per project just means agents stepping on each other.
• Also :: Smaller teams, slowly falling junior/intern hiring, and agentic infra as its own discipline.

What did not change

• Team is still the unit of work :: Anthropic and OpenAI keep two-pizza teams that own software and on-call. "Two humans and a bunch of agents" is a one-off project, not a team.
• Planning :: Still essential for complex infra work, team alignment and customers. Building the wrong thing is waste even when building is cheap.
• Testing :: Jarred Sumner (Bun): AI writes almost all code and tests, but test effort still matches code effort. Tests are how you trust code; formal verification is rising.
• Non-engineers don't ship to prod :: They build sites and report bugs via Slack bots. Nobody Gergely asked has non-engineers shipping production code.
• Old patterns rediscovered :: Matt Pocock found that vocabulary from classic books ("tracer bullets", deep/shallow modules) makes agents write better code.

What broke

• Code volume :: GitHub PRs/commits/repos up ~5x over three years and accelerating, with PRs 2x in the last two months. Hence the GitHub outages.
• Code review :: "Zombie code reviews": everyone fakes it and rubber-stamps LGTM. Ramp now trusts engineers to skip review on non-critical code. More will follow.
• Quality :: Visible paper cuts everywhere (Spotify, Substack…).
• Infra capacity :: Beyond GPU and memory shortages there is now a CPU shortage. Server lead time went from 1–2 weeks to 6 months. Some regions take no reservations, and companies prepay for capacity arriving in December.
• Focus :: Constant context switching between waiting agents. Parallel work became the baseline expectation, so people feel they have more work than pre-AI.
• Engineering leadership :: Many leaders are leaving. Founder-mode pressure (fire half the team, 2x revenue), failing startups, orgs too slow on AI (a career dead end), fewer middle managers, rising fractional-CTO work, AI labs outbidding on comp (Anthropic hiring ex-CTOs as ICs), easy VC for hands-on leaders, and burnout.

What's next

• Cloud agents and harnesses :: Ramp's in-house cloud harness beats off-the-shelf tools: the cloud is less limited than laptops, has better frontend tooling, and doubles as remote dev environments. Dima (Linear) flipped from local-dev fan to cloud believer.
• We will stop reading code :: Even skeptic Charity Majors says shipping code you haven't read is a "when, not if", as ops and QA always have.
• New internal infra :: Evals in CI/CD, iterated agentic factories, agents in deploys, observability and incident management.
• Refactoring wave :: Cheap migrations and rewrites, plus a low-risk way to practice agentic skills.
• AI fluency as a hiring baseline :: Some startups screen for "AI positivity".
• AI maturity :: Lara Hogan: mature orgs start from a business outcome and build team- or company-level agentic systems that cut handoffs, not just individual automations.
• Domain depth wins :: Titus Winters: success = intelligence (how), wisdom (what) and charisma (convincing). With "how" commoditized, wisdom and charisma matter more.

Advice for engineering leaders

Stay hands-on and ship to production. Will Larson shipped more code in 15 months than in the previous 5–10 years. Build AI infra for your team and company, use AI to remove friction, and don't outsource your learning. Accept less people management for now. Peter Mattis went back to hands-on and says he learned more in the past year than in the previous five. Remember why you got into tech: building things, now faster.