How AI is reshaping software engineering and the startup economy

Executive overview

AI coding tools have already transformed how engineers work — at OpenAI, 95% of engineers use Codex daily and virtually all code is AI-authored. Engineers are no longer writing code; they are steering fleets of agents, managing context, and reviewing outputs.

The shift goes beyond tooling. Agents will soon handle multi-hour tasks autonomously, business process automation is a vastly underrated opportunity, and a coming startup boom may produce millions of small, high-value businesses rather than a handful of unicorns.

The models will eat your scaffolding for breakfast — build for where they're going, not where they are today.

AI and code at OpenAI today

• 95% of engineers use Codex daily; 100% of PRs are reviewed by Codex
• Engineers using Codex open ~70% more PRs, and the gap is widening
• Code reviews have shrunk from 10–15 minutes to 2–3 minutes with AI suggestions
• CI automation (lint fixes, test reruns) is largely handled by Codex
• Codex reviewing its own PRs is standard; human attention has dropped from 100% to ~30% per PR

The engineer as sorcerer

• The role has shifted from writing code to managing 10–20 parallel agent threads simultaneously
• Engineers are now tech leads by default — steering, checking, and unblocking agents
• The SICP "wizard book" metaphor from 1980 has literally come true: programming is now incantations
• The Sorcerer's Apprentice risk is real — agents need supervision; unchecked agents cause chaos
• A 100% Codex-written codebase experiment reveals the core constraint: agents fail when context is underspecified
• Fix: encode tribal knowledge into the repo via code comments, .md files, and structured documentation

The role of the engineering manager

• Managers are less directly affected than ICs, but the leverage gap between performers is widening
• Spend more than 50% of management time with top 10% performers — they compound fastest with AI
• The mythical Man-Month surgeon analogy applies: the manager's job is to have the scalpel ready before the surgeon asks
• AI can surface active blockers by querying Slack, Notion, and GitHub — and could predict next month's blockers
• Managers will likely oversee larger teams as AI reduces the coordination overhead of each direct report

Why most AI deployments have negative ROI

• Companies outside tech are not power users — they ask basic questions and don't push the models
• The failure pattern: top-down mandate with no bottoms-up adoption
• Workers know they're supposed to use AI but have no peers to learn from and no one modelling best practices
• Fix: create a dedicated internal tiger team of technically adjacent (not necessarily engineering) enthusiasts
• Let them run hackathons, build workflows, do knowledge-sharing sessions — then spread what works
• Top-down buy-in is necessary but not sufficient; the bottoms-up energy is what actually moves adoption

Building on the API without being squashed

• The market is so large that OpenAI building in a space does not eliminate the opportunity
• Cursor thrived in the most competitive coding space because users genuinely loved it
• Failed startups almost never failed because a lab competed — they failed because the product didn't resonate
• OpenAI views itself as an ecosystem platform: every model released internally also ships to the API
• 800 million weekly active users on ChatGPT create massive distribution for third-party builders via the GPT app store
• OpenAI's mission ("spread benefits to all of humanity") requires a platform strategy — they cannot reach every use case alone

Don't listen too literally to customers in AI

• The field changes so fast that customer requests reflect a local maximum, not the optimal path
• "The models will eat your scaffolding for breakfast" — vector stores, agent frameworks, and file-based context management are all transient
• What was essential scaffolding in 2022 (vector stores, RAG pipelines) is now often unnecessary overhead
• Build for where the model capability will be in 12–18 months, not where it is today
• Products built slightly ahead of capability unlock dramatically when the next model ships

Second-order effects of the one-person startup

• The one-person billion-dollar startup implies it is now easy to start any startup, not just unicorns
• Likely outcome: an explosion of $10M–$100M micro-companies that are excellent for founders but poor for VC returns
• A golden age of B2B SaaS — hundreds of small startups building bespoke software for other small startups
• VC return dynamics may compress as venture-scale companies become rarer relative to lifestyle businesses
• Distribution and audience become more valuable as the number of products competing for attention explodes

Business process automation is underrated

• Most economic activity is repeatable, deterministic, procedure-driven work — not open-ended knowledge work
• Software engineering is the exception; support, operations, and finance are the rule
• AI is well-suited to high-determinism, rules-based processes integrated with enterprise data systems
• This opportunity is rarely discussed in tech circles because it is outside the Silicon Valley frame of reference
• Impact on business process jobs may ultimately exceed the impact on software engineering roles

Where the platform is heading in 12–18 months

• Agents will reliably handle multi-hour tasks; the SWE-bench task-length curve is rising steeply
• Products will need to be redesigned around longer-running, less interactive agent workflows
• Audio and speech models are significantly underrated — most business is conducted verbally
• Native multimodal audio will unlock enterprise use cases in support, operations, and services
• The API stack: Responses API (low-level) → Agents SDK (orchestration) → Agent Kit + Widgets (UI) → Evals API (testing)