The state of the tech industry in 2026 - LDX3 New York

Gergely Orosz surveys how coding agents are changing engineering work, drawing on visits and conversations with AI labs, startups, and larger technology companies. His central tension: generating more code is becoming easier, while coordinating work, validating changes, and deciding what to build remain engineering responsibilities.

Evidence basis: Summarized from English auto-generated YouTube captions, which contain transcription errors, without an original-audio check. Company practices and platform statistics are Orosz's reporting, not independently verified here. His observations emphasize AI-intensive companies; broad adoption claims are not a representative industry survey, and predictions below remain forecasts.

More agents change both delivery and coordination

• Parallel agents replace a single stream of implementation work. Orosz describes engineers running five to ten agent sessions, sometimes using separate Git worktrees—working copies that let tasks proceed independently. Companies also build internal harnesses, the software around an agent that connects its tools and data, and initiate work through Slack. His OpenAI example includes agents monitoring production and proposing performance changes, while humans still often review critical, large-impact pull requests. It is an evolving system, not fully autonomous delivery. 00:53–02:47(opens in a new tab), 04:31–07:23(opens in a new tab)
• Migration accounts suggest cheaper large changes, not instant rewrites. He reports an ongoing OpenAI Python-to-Rust API migration reaching roughly 90% in four to five months, Airbnb's testing migration taking six weeks, and Uber's JUnit migration taking four months. He qualifies Asana's two-week testing migration with years of prior work. These are reported project examples, not controlled comparisons proving a universal speedup. 06:21–08:13(opens in a new tab)
• Smaller projects do not eliminate durable teams. An Anthropic account describes limiting individual projects to one or two engineers because each brings several agents whose changes can interfere. Orosz separately stresses that teams still own software, on-call coverage, quality standards, and continuity when people take leave. Complex infrastructure work still needs planning and coordination; building the wrong thing faster remains waste. 11:02–13:57(opens in a new tab)

Validation and attention do not scale automatically

• Tests remain necessary even when agents write them. Orosz recounts the Bun team's roughly unchanged balance between time spent on production code and tests, although AI now writes much of both. He contrasts this need for validation with reports of overloaded reviewers approving changes with little scrutiny. Ramp's selective removal of reviews for non-critical code is one response, not evidence that every team can safely abandon review. His broader claim of declining software reliability is an observation, not a measured causal finding about AI. 13:57–14:45(opens in a new tab), 16:44–18:37(opens in a new tab)
• Non-engineers initiating work is different from owning production delivery. In the companies he asked, non-engineers build websites or flag bugs through Slack, where a bot may open a pull request; engineers still handle production code. This is the boundary he observed, not proof that non-engineers never ship elsewhere. 14:45–15:48(opens in a new tab)
• Parallelism can raise expectations faster than it reduces workload. A Linear engineer describes constant context switching as agents wait for answers. What initially felt like a productivity advantage became the expected pace, leaving him feeling busier than before AI. Orosz also reports engineering leaders facing unrealistic staffing or revenue expectations, slower-moving employers, and burnout; he does not attribute every departure to AI. 18:37–20:34(opens in a new tab), 20:34–22:35(opens in a new tab), 22:35–23:38(opens in a new tab)

His forecast favors shared infrastructure and hands-on leadership

• Cloud agents solve local capacity problems but need available infrastructure. Ramp's stated reasons include escaping laptop limits, providing better frontend tooling, and establishing remote development environments. A Linear engineer similarly values agents that do not occupy his machine. Orosz expects wider adoption, while warning that CPU capacity constraints can complicate it. 22:35–24:30(opens in a new tab)
• Model choice and routing can change the cost curve. He reports companies investing in open models and smarter model routing to lower per-token costs and keep spending flat. This is a reported operational response to rising bills, not a guarantee that every team will save the same amount. 09:07–11:02(opens in a new tab)
• Improve shared workflows, not only individual automation. Orosz predicts more agent involvement in deployment, observability, and incident management, and forecasts engineers shipping code without personally reading it. These are forecasts, not a demonstrated replacement for understanding or verification. Drawing on Lara Hogan, he recommends starting with a business outcome and working backward to reduce handoffs and friction across teams. His advice to leaders is to stay hands-on, build enabling infrastructure, and use AI to learn rather than outsource their learning. 23:38–27:13(opens in a new tab), 27:13–28:52(opens in a new tab)

Full video: https://www.youtube.com/watch?v=Ru99FGJ_yuE(opens in a new tab)