Lessons from Alex Embiricos

Lessons from Alex Embiricos

Alexander Embiricos leads development for OpenAI's coding tools as Head of Codex. He previously co-founded Multi, a virtual office for remote engineers. This collection gathers his thinking on software development, autonomous agents, and the limits of human-computer interaction.

Part 1: The Bottleneck to AGI

1. On AGI limits: "The current underappreciated limiting factor to AGI is human typing speed or human multi-tasking speed on writing prompts." — Source: [Business Insider]
2. On human constraints: "As models become more capable, the bottleneck shifts away from the AI's ability to write code toward human validation and judgment." — Source: [The Twenty Minute VC]
3. On prompt throughput: "You realize quickly that the system can generate solutions faster than you can read them, which completely changes where the latency lies in the development loop." — Source: [How I AI]
4. On evaluating output: "We spend more time thinking about how a developer will verify a piece of code than how the model will generate it." — Source: [AI + a16z]
5. On the verification tax: "Generating the code is cheap. Figuring out if the code does what you actually wanted is where all the time goes now." — Source: [The Twenty Minute VC]
6. On human multi-tasking: "Developers max out on how many parallel AI tasks they can supervise long before the models run out of capacity to serve them." — Source: [How I AI]
7. On shifting focus: "Our engineering problem has shifted from making the model smarter to making the human faster at directing the smart model." — Source: [Business Insider]
8. On judgment as a skill: "The primary skill of a senior engineer is moving from syntax generation to rapid architectural judgment." — Source: [AI + a16z]
9. On the reading bottleneck: "If an AI writes ten thousand lines of code in a minute, you still have to read and trust those ten thousand lines before shipping." — Source: [The Twenty Minute VC]

Part 2: The Future of Coding Agents

1. On moving past pair programming: "The future of coding is moving from simple pair programming to delegation, where AI acts as a teammate capable of managing complex, end-to-end tasks." — Source: [How I AI]
2. On agent autonomy: "We are transitioning from tools that autocomplete your current line to systems that execute your intent across an entire repository." — Source: [The Twenty Minute VC]
3. On competitive tools: "The space between Codex, Cursor, and Claude Code is less about who has the rawest intelligence and more about who builds the best integration into the developer's daily habits." — Source: [The Twenty Minute VC]
4. On delegation vs dictation: "You want to assign a ticket to the agent, not dictate the keystrokes it should type." — Source: [How I AI]
5. On tool convergence: "Eventually, the specialized coding tools will converge into a more universal agentic capability that understands context implicitly." — Source: [AI + a16z]
6. On context windows: "Feeding the whole repository into context isn't enough; the agent needs to know which files actually matter for the specific bug you are fixing." — Source: [How I AI]
7. On pull request success rates: "When we look at high PR success rates, it usually comes down to how well the agent breaks down the problem before writing any syntax." — Source: [AI + a16z]
8. On the lifecycle of code: "Agents need to understand that writing the code is step one. Step two is getting it past CI, and step three is deploying it safely." — Source: [The Twenty Minute VC]
9. On agent reliability: "A developer will forgive a tool that doesn't know the answer, but they will quickly abandon a tool that confidently breaks their build." — Source: [AI + a16z]
10. On standardizing intent: "The interface for coding will eventually look much closer to plain English specifications than specialized syntax." — Source: [How I AI]

Part 3: The Three Phases of AI Agents

1. On phase definitions: "We see agent development happening in three distinct phases: coding assistance, general computer use, and fully productized workflows." — Source: [The Twenty Minute VC]
2. On the coding phase: "Coding was the natural first phase because code has a built-in verification mechanism. It either compiles and passes tests, or it doesn't." — Source: [AI + a16z]
3. On computer use: "The second phase is giving the AI the ability to use the computer exactly like a human does, navigating interfaces that lack APIs." — Source: [The Twenty Minute VC]
4. On workflow productization: "Phase three is when these agents stop being raw tools and become packaged, reliable workflows that run completely in the background." — Source: [How I AI]
5. On transitioning phases: "Moving from phase one to phase two requires a leap in how the model understands visual context and state changes on a screen." — Source: [The Twenty Minute VC]
6. On API reliance: "You can't wait for every tool to have a perfect API. Agents need to interact with the messy, GUI-driven reality of modern software." — Source: [AI + a16z]
7. On background execution: "The ultimate goal is an agent that takes a request on Friday and has the workflow completed, tested, and staged by Monday morning." — Source: [How I AI]
8. On phase overlap: "These phases are not strictly sequential. We are already seeing early signs of productized workflows even as we perfect basic coding." — Source: [The Twenty Minute VC]
9. On testing agents: "Validating an agent that uses a GUI is exponentially harder than validating an agent that outputs plain text code." — Source: [AI + a16z]

Part 4: Redefining Product Management

1. On eliminating modes: "We're really thinking about how we get this to ChatGPT in such a way that you're not thinking about modes. No modes for us. That is the goal." — Source: [AI Leaders Council]
2. On seamless integration: "The user should not have to decide which version of the model to use for a specific task. The system should infer it." — Source: [How I AI]
3. On model personality: "Listening to power users showed us that the personality of the output matters. Some users want verbose explanations; others want terse, concise answers." — Source: [AI Leaders Council]
4. On the PM role: "The role of a product manager in an AI-native world is largely undefined right now. It is shifting from defining strict user flows to defining acceptable bounds of model behavior." — Source: [The Twenty Minute VC]
5. On user feedback: "A single tweet from a power user explaining their friction with verbosity changed how we thought about default model behavior." — Source: [AI Leaders Council]
6. On deterministic design: "We are moving away from deterministic state machines. Product management now requires a deep comfort with probabilistic outcomes." — Source: [The Twenty Minute VC]
7. On feature bloat: "If you add a toggle for every possible user preference, you have failed at product design. The AI should adapt without configuration." — Source: [How I AI]
8. On building trust: "Trust is built when the product consistently respects the user's implicit preferences without needing explicit instructions every time." — Source: [AI Leaders Council]
9. On rapid iteration: "The feedback loop for AI products is measured in hours. You ship a change, and the community immediately tests its edge cases." — Source: [The Twenty Minute VC]
10. On intuitive interfaces: "The best interface for an AI agent is one that disappears entirely, leaving only the task to be done." — Source: [How I AI]

Part 5: Engineering Workflows in the AI Era

1. On spec-first development: "Writing a structured plan before generating code is the single highest leverage activity a developer can do when working with AI." — Source: [How I AI]
2. On structured planning: "We encourage using a Plans.md file to align the agent on the architectural goals before it writes the first line of logic." — Source: [How I AI]
3. On parallel development: "By utilizing Git worktrees, a developer can have an agent exploring one solution branch while they manually debug another." — Source: [How I AI]
4. On internal tooling: "We rely heavily on automated code review internally to catch the subtle logic errors that large language models sometimes introduce." — Source: [AI + a16z]
5. On speed of execution: "Using these agentic workflows, we were able to build the Sora Android application in just 28 days." — Source: [How I AI]
6. On code review changes: "The bottleneck has moved from writing the pull request to reviewing the pull request. We need AI to help review the code that AI wrote." — Source: [The Twenty Minute VC]
7. On debugging: "Agents are often better at writing new code than they are at navigating a sprawling legacy codebase to find a tiny state bug." — Source: [AI + a16z]
8. On modularity: "AI tools force you to write more modular code, because smaller, isolated functions are much easier for a model to reason about." — Source: [How I AI]
9. On technical debt: "If you aren't careful, an agent can generate technical debt faster than any human team ever could. Architectural oversight is critical." — Source: [The Twenty Minute VC]

Part 6: Building Remote & Hybrid Tools

1. On the virtual office: "We built a highly crafted amazing tool to use that makes working remotely feel a little bit more like you’re sitting next to each other." — Source: [Code Story]
2. On presentation tools: "A lot of remote tools were built around presentations rather than active problem-solving and fluid collaboration." — Source: [Code Story]
3. On spontaneous interaction: "The hardest thing to replicate in a remote environment is the casual, low-friction tap on the shoulder." — Source: [IT Career Energizer]
4. On team trust: "Trust in remote teams is built through frequent, lightweight interactions, rather than weekly formal status updates." — Source: [TEAMES & CO]
5. On technical collaboration: "We rebranded Remotion to Multi because we realized technical teams needed dedicated spaces for pair programming and deep collaborative work." — Source: [Multi Blog]
6. On synchronous work: "Asynchronous work is great for deep focus, but synchronous work is where actual alignment and creative problem solving happens." — Source: [Code Story]
7. On tool friction: "If it takes three clicks and a calendar invite to start a conversation, the conversation simply won't happen." — Source: [IT Career Energizer]
8. On hybrid environments: "Hybrid is often harder than fully remote because it creates two classes of citizens on the team if the tooling isn't perfect." — Source: [TEAMES & CO]
9. On screen sharing: "Screen sharing shouldn't feel like a broadcast; it should feel like two people pointing at the same whiteboard." — Source: [Multi Blog]
10. On developer needs: "Engineers specifically need tools that stay out of the way of their IDE while still allowing high-bandwidth communication." — Source: [Multi Blog]

Part 7: Customer Discovery & Founder-Market Fit

1. On choosing a problem: "We ended up choosing Remotion partially on the basis of those interviews, and then partially also based off our own gut feel of which problem we had maybe better founder-market fit for." — Source: [Code Story]
2. On founder passion: "We were really passionate about this idea of having very fluid conversation even though you were on a remote team." — Source: [Code Story]
3. On user interviews: "Customer discovery is less about asking people what they want and more about observing where they currently experience friction." — Source: [IT Career Energizer]
4. On pivoting: "The decision to pivot to Multi was driven by recognizing that our most engaged users were engineering teams trying to pair program." — Source: [Multi Blog]
5. On trusting intuition: "Data will tell you what is happening, but founder intuition is required to figure out why it is happening and what to build next." — Source: [Code Story]
6. On niche markets: "It is always better to build something that a small group of people love rather than something a large group of people merely tolerate." — Source: [IT Career Energizer]
7. On validating ideas: "You validate an idea by seeing if someone will actually change their daily habit to use your prototype, not by asking if they like the concept." — Source: [Code Story]
8. On target audiences: "Focusing entirely on technical teams allowed us to build features that a general-purpose communication tool could never justify." — Source: [Multi Blog]
9. On early feedback: "The harshest feedback from early users is usually the most valuable signal you can get." — Source: [IT Career Energizer]

Part 8: Career Advice & Leadership

1. On choosing customers: "In your job, make sure that the people you deal with—your customer base—are the kinds of people that you enjoy talking to." — Source: [IT Career Energizer]
2. On empowering others: "Always do your best to empower others so that they can solve problems for themselves." — Source: [IT Career Energizer]
3. On continuous learning: "The landscape of software development is changing so fast that the ability to unlearn old habits is now more important than learning new ones." — Source: [The Twenty Minute VC]
4. On early career choices: "Optimize for learning rate rather than title in the first five years of your career." — Source: [Code Story]
5. On failure: "A failed startup teaches you more about product distribution and team dynamics than a successful stint at a massive corporation." — Source: [IT Career Energizer]
6. On leadership communication: "Clear writing is a prerequisite for clear leadership, especially in distributed organizations." — Source: [TEAMES & CO]
7. On hiring engineers: "We look for engineers who are product-minded and care about the user experience alongside the technical elegance of the solution." — Source: [Multi Blog]
8. On managing energy: "Burnout usually isn't about working too many hours; it is about working on things that drain your energy instead of replenishing it." — Source: [Code Story]
9. On technological shifts: "When a platform shift happens, the worst thing you can do is hold onto the assumptions of the previous era." — Source: [The Twenty Minute VC]