20VC: Codex vs Claude Code vs Cursor: Who Wins, Who Loses | Will All Coding Be Automated - Do We Need PMs | The Real Bottleneck to AGI | The Three Phases of Agents and What You Need to Know with Alex Embiricos, Head of Codex at OpenAI

What It Covers

Alex Embiricos, Head of Codex at OpenAI, maps the three phases of coding agents—from interactive pair programming to cloud delegation to full workflow automation—while addressing whether Cursor will lose half its revenue, why human validation bottlenecks AGI more than compute, and where SaaS companies remain defensible against model providers.

Key Questions Answered

• •Three Phases of Agent Adoption: Coding agents evolve through three distinct stages: first, specialized coding tools where LLMs already excel; second, general-purpose agents accessible to any builder via flexible interfaces like the Codex app; third, productized vertical features that work out-of-the-box. Teams currently in phase two should resist over-specifying workflows before users develop fluency with the underlying tools, or adoption stalls entirely.
• •Human Validation as the AGI Bottleneck: The primary constraint on AI deployment is not model capability, compute, or architecture—it is the human effort required to prompt, manage, and validate agent output. Most users interact with AI roughly 30 times daily, but frictionless AI should assist tens of thousands of times per day. Removing the need for users to recognize when AI can help—through proactive, context-aware agents—is the core product challenge to solve.
• •Delegation Over Pairing as the New Workflow: Since GPT-4.5 Codex launched in December, OpenAI engineers largely stopped opening IDEs. The shift moved from pair-programming—where humans stay at the keyboard—to full task delegation: writing a spec, reviewing the agent's plan, then letting it execute independently. The Codex app was built specifically around this delegation model, removing text editing entirely to reinforce the behavioral change.
• •Plan Review Replaces Code Review: As agents write the majority of code, reviewing the agent's proposed plan before execution becomes more valuable than reviewing the resulting code. Codex now includes a prominent plan mode where the agent proposes its approach and asks clarifying questions before starting—mirroring how a new hire would present a request-for-comments. Additionally, Codex automatically reviews nearly all code pushed to OpenAI repos, trained to produce high-signal, low-false-positive feedback.
• •SaaS Defensibility Depends on Two Assets: SaaS companies remain defensible if they own either a direct human relationship or a critical system of record—ideally both. Companies acting purely as integration glue layers without owning either face the highest displacement risk. Embiricos specifically flags customer support as a category OpenAI will enter, while arguing that companies in gnarly, relationship-dense markets—such as fintech with complex banking integrations—are structurally harder for model providers to displace.
• •Open Standards as Competitive Strategy: Codex pursues retention through openness rather than lock-in: the core harness is open source, and OpenAI initiated the agents.md and .agents/skills standards so any agent can read configuration files. Stickiness increases naturally as agents connect to enterprise systems—Sentry, Google Docs, internal tools—because those integrations require security, permissioning, and trust decisions that enterprises will not repeat. Winning the integration layer early creates durable retention without artificial switching costs.