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

Summary of main arguments and developments (Tech industry in 2026)

Gergely argues that the tech industry has undergone rapid, structural change over the last ~12 months, driven primarily by AI-assisted software development. He frames the update around what has changed, what has stopped working (“broke”), and what to expect next, based on visits and interviews with major AI and tech companies.

---

What has changed

1. Code is increasingly generated by AI (and nearly replaces hand-coding)

   • Engineers now mostly rely on AI models to write code rather than writing it manually.
   • Internal/startup metric trendlines suggest AI-generated code is approaching near-total adoption.

2. Parallel “multi-agent” workflows have become normal

   • Engineers commonly run multiple agent sessions at once (multiple IDE tabs/agents/agent windows).
   • This increases throughput but also raises cognitive load (e.g., keeping many agent copies and working across multiple code branches).

3. Traditional IDEs are declining

   • Agent-based interfaces and tools are replacing heavyweight IDE workflows.
   • He claims forks of VS Code are moving toward “agent-first” experiences (e.g., Cursor shifting to agent-only interfaces; JetBrains refocusing toward agent-like tooling).
   • IDEs are described as shrinking in importance—still useful in corporate contexts, but less central to modern development.

4. Companies are building internal AI “platforms” and Slack-centered development

   • Development begins inside tools like Slack, where AI actions are triggered (e.g., “implement this”).
   • Many organizations integrate AI agents directly into their internal systems and workflows.

5. Agent “software factories” are emerging

   • He highlights OpenAI’s “Agent Factory” concept (and related monitoring/performance improvements), where agents act on production signals and generate pull requests automatically.
   • Automation is not fully complete, but it signals a shift toward continuous agent-driven engineering.

6. Migrations, rewrites, and refactoring are faster

   • AI makes large engineering changes cheaper and dramatically quicker, with examples including:
     • Anthropic’s migration from Zig to Rust (reported as rapid due to AI assistance)
     • OpenAI’s API migration (Python to Rust), reaching ~90% quickly and expected to finish soon
     • Airbnb test migration reportedly done in weeks; Asana in ~2 weeks
     • Uber’s JUnit 4 to 5 migration, cited as accelerated

7. Agents are generating more work than humans

   • Using data from Linear and GitHub, he claims agent-generated tickets and pull requests are rising sharply.
   • In some cases, agent-generated output exceeds human-generated output.

8. AI “skills” are mainstream

   • He references monitoring indicating most users are now using AI skills in tools/platforms (up from far lower adoption earlier in the year).

9. Cost pressure led to optimization

   • After cost overruns (e.g., Uber’s narrative of exhausting its AI budget), companies reduce cost per token by:
     • investing in open models
     • using smarter routing
     • optimizing infrastructure
   • He claims cost per token reductions on the order of ~50% while maintaining spend levels.

10. Engineering specializations are fading

    • LLM assistance reduces the value of narrow role boundaries (e.g., mobile/frontend engineers handling broader tasks).
    • Projects increasingly involve very small teams (often 1–2 engineers), since each engineer already runs multiple agents.

11. Teams are shrinking; juniors and interns are hiring down

    • Hiring signals suggest fewer junior roles and continued decline in intern hiring.
    • “Agent infrastructure” is described as becoming a distinct discipline.

---

What hasn’t changed (according to the speaker)

1. Teams remain the core unit of work

   • Even with lots of agents, organizations still organize around real teams and accountability.

2. Planning and alignment still matter

   • For complex infrastructure, planning resembles older patterns—speed alone isn’t justification for wasted products.

3. Testing/validation still consumes substantial effort

   • Even when AI writes most code, teams must still build and run tests (with AI assistance).
   • Tests remain critical for trust/validation, with continued need for formal verification in some contexts.

4. Non-engineers still don’t “ship production code”

   • He rejects claims that non-engineers can directly submit production-ready code.
   • Non-engineers typically report bugs, tag engineering teams, or provide input; AI may draft changes, but releases still require engineering responsibility.

5. Old design principles are being re-learned for AI-directed development

   • Concepts such as “tracer bullets” and modularity reappear as effective techniques when prompting AI.

---

What “broke” (pain points and dysfunctions)

1. Code volume skyrocketed, but quality/review practices broke

   • GitHub data trends show major increases in agent-generated pull requests and overall PR volume.
   • As a result, code reviews become ineffective, turning into “zombie reviews” (formal checks without meaningful scrutiny).
   • Some companies waive reviews for low-risk changes or eliminate them selectively.

2. Software reliability issues are increasingly visible

   • He claims “little flaws” show up across public services, implying QA/review throughput hasn’t scaled with agent-driven output.

3. Compute/infrastructure shortages got worse

   • Beyond GPU shortages and memory constraints, he highlights an unexpected CPU shortage affecting cloud providers.
   • Delivery delays are described as severe (e.g., CPU/server ordering taking up to ~6 months, with some regions not accepting new orders).

4. Productivity hit “context switching” limits

   • Developers juggle many concurrent agent tasks and constantly wait on or coordinate with other agents.
   • Early AI gains (“do in hours what took a day”) evolve into a new norm: parallel work increases total workload and reduces real perceived productivity.

5. Technical leadership instability

   • He claims technical leaders are burning out or leaving due to:
     • unrealistic AI-driven expectations from leadership/founders (e.g., layoffs, doubling revenue)
     • companies failing to implement AI workflows quickly enough
     • startups collapsing or slowing down
     • smaller teams reducing management layers
     • better compensation and opportunities for AI-experienced executives at AI startups
     • role confusion as “factional CTOs” (part-time/contract leadership) become more popular

---

What to expect next

1. More cloud-based agent development and automation

   • He contrasts cloud vs on-prem approaches:
     • cloud development reduces local overload and can be faster/easier
     • local development is still valued by some, but cloud is presented as the direction due to capability and scaling
   • He notes potential security tradeoffs but frames them as manageable.

2. Reading/verifying code becomes less central

   • He attributes a “don’t read code” direction to skepticism leaders (Charity Majors is cited).
   • The expectation is that QA/ops-style trust mechanisms and evaluations will be embedded into pipelines.

3. Evaluations (evals) become part of CI/CD

   • Continuous validation is expected to become standard alongside automated deployments.

4. Agent factories integrate into deployment/monitoring/incident management

   • Agents move beyond writing code into the full operational lifecycle.

5. Another wave of refactoring/migrations/code rewrites

   • If AI makes change cheaper, organizations will repeatedly restructure systems.

6. AI fluency becomes a hiring baseline

   • A CTO screening candidates for positive AI attitude is referenced.
   • He also claims leaders should expect maturity-building around agent systems and infrastructure.

7. Organizational strategy shifts: outcome-driven agent systems

   • Advice cited: top-performing organizations start from business outcomes and build agent systems to reduce handoffs and friction—rather than doing simple individual automation.

8. Wisdom and judgment become more important than raw “how-to”

   • As tool-assisted capability (“intelligence”) becomes widely accessible, senior success will depend more on wisdom and judgment: choosing the right direction and aligning people.

9. Practical career and leadership guidance

   • Leaders should stay hands-on, build AI infrastructure, use AI to reduce friction, and maintain ownership rather than delegating responsibility.
   • He emphasizes studying and continued practice with AI tools.

---

Presenters / contributors mentioned

• Gergely (presenter; writes for Pragmatic Engineer and hosts a software development podcast)
• Boris Cherny
• Dima Zachev
• Peter Mattis (co-founder and CEO, Cockroach Labs)
• Kent Beck
• Solomon Chachari
• Kathleen Less
• Jared Summer (Bun team)
• Matt Pocock (creator/educator; references work on “tracer bullets”)
• Will Larson (chief product and technology officer)
• Caitlin (speaker referenced in Anthropic discussion; first name only)
• Titus Winters
• Laura Tacho
• Charity Majors