## Summary

In his keynote presentation at LDX3 New York, Gergely Orosz synthesizes observations from on-the-ground visits to leading AI laboratories and technology companies—including OpenAI, Anthropic, Cursor, Ramp, and Linear—to map the structural transformation of the software engineering landscape. Framing his discourse across four thematic pillars—what changed, what remained constant, what broke, and what lies ahead—Orosz examines how the shift to agentic engineering has fundamentally restructured development workflows. He illustrates how raw code generation has shifted almost entirely to AI models, traditional heavyweight IDEs have yielded to cloud-based agent harnesses, multi-year migrations are executed in weeks, and project teams are consolidating into one- or two-person squads managing parallel AI agents.

To counterbalance exaggerated industry narratives, Orosz systematically identifies enduring software engineering fundamentals, asserting that collaborative human teams remain the indispensable unit of ownership, rigorous planning remains critical to avoid architectural missteps, and testing demands equal effort to code synthesis. He debunks claims that non-engineers are shipping production code and highlights the rediscovery of classic modular design paradigms used to steer agent reasoning. However, this rapid operational acceleration has introduced severe friction across the ecosystem: pull request volumes have overwhelmed peer review into performative 'zombie code reviews', software reliability has degraded with continuous small regressions, supply chain bottlenecks have expanded into severe CPU hardware shortages, and developer context switching has heightened cognitive fatigue.

Progressing to future trajectories, the narrative pivots to strategic guidance for engineering leadership navigating organizational restructuring and executive burnout. Orosz argues that development environments will shift decisively to the cloud and that organizations will transition toward shipping verified code without manual line-by-line inspection via automated evaluation pipelines. Emphasizing that pure execution 'intelligence' is becoming commoditized, he contends that architectural 'wisdom' and persuasive 'charisma' are the defining differentiators for engineers. The keynote culminates in an urgent call for engineering leaders to remain hands-on builders, implement organization-level agentic infrastructure, and rediscover the core joy of building software.

## Topics

**Introduction and Scope of Industry Research**
- Gergely outlines his keynote based on firsthand visits and conversations inside OpenAI, Anthropic, Cursor, Ramp, Uber, and Linear.
- The presentation adopts a four-part framework: what changed, what has not changed, what broke, and what comes next.

**What Changed: The Agentic Engineering Shift**
- The majority of production code is now AI-generated via models such as Opus 4.5 and GPT-5.2 rather than manually written.
- Hardcore engineers routinely manage 5 to 10 parallel agent sessions across concurrent worktrees.
- Heavyweight IDEs are rapidly declining into legacy products as teams adopt agent harnesses like Cursor, JetBrains Air, and OpenAI Codex.
- Startups and large tech companies are building proprietary internal agent harnesses integrated directly into Slack workflows.
- OpenAI and others have deployed autonomous agentic software factories and performance optimization pipelines.
- Large-scale codebase migrations (e.g., Python to Rust, test framework rewrites) that once took years are now completed in weeks or months.
- Engineering specializations are flattening, individual project staffing has shrunk to 1–2 engineers, and junior hiring continues to slow down.

**What Has Not Changed in Software Engineering**
- Two-pizza teams remain the vital organizational unit for accountability, on-call ownership, and collaboration.
- Architectural planning and customer alignment remain essential to avoid rapidly building the wrong systems.
- Testing and verification continue to consume roughly equal time and focus as code generation.
- Non-engineers are still not deploying production code independently despite low-code and AI hype.
- Classic software architecture concepts (e.g., deep modules, tracer bullets) are being rediscovered to prompt AI agents more effectively.

**What Broke: Reviews, Quality, Infrastructure, and Leadership**
- Massive code volume surges have rendered traditional peer code review obsolete, giving rise to rubber-stamped 'zombie reviews'.
- Product quality and application reliability have deteriorated under a steady accumulation of agent-generated bugs and paper cuts.
- Hardware constraints have expanded beyond GPUs to acute CPU server procurement shortages with multi-month lead times.
- Relentless context switching across multiple agent threads is straining developer focus and productivity.
- Engineering leaders face severe burnout, elevated turnover, and shifting corporate expectations amid downsizing and lagging enterprise AI adoption.

**What's Next and Strategic Advice for Leaders**
- Development environments and agent harness platforms are migrating entirely toward cloud infrastructure.
- Engineering organizations will shift toward shipping code without manual line-by-line reading by leaning heavily on automated eval suites.
- Hiring criteria increasingly mandate AI fluency and active positivity toward agentic tooling.
- Maturing organizations are shifting focus from isolated task automations to cohesive team- and company-level agent architectures.
- Because technical execution intelligence has become cheap and accessible, architectural wisdom and persuasive charisma are the new core differentiators.
- Leaders must remain hands-on builders, deploying internal AI infrastructure and actively shipping to production.