The Future of Software

01Theme 1: Engineers Are Becoming Sorcerers

The most striking data point from the conversation: at OpenAI, 95% of engineers use Codex daily and 100% of PRs are reviewed by Codex. Engineers who use Codex more open 70% more PRs — and the gap is widening. IC engineers are essentially becoming tech leads, managing fleets of 10-20 parallel AI agent threads simultaneously.

Sherwin draws on SICP (Structure and Interpretation of Computer Programs), the MIT textbook from 1980 that described programming as sorcery — languages as incantations, engineers as wizards casting spells. That metaphor, he argues, is now literally playing out: “It literally feels like we're wizards now... we're closer to having it feel like this magical experience where we're casting all these spells and having software do all these things for you.”

The Sorcerer's Apprentice from Fantasia is the sharper analogy. When engineers dispatch 20 Codex threads, there's real skill and seniority required — you don't want the brooms going off the rails. But it's extremely high leverage.

02Theme 2: The One-Person Billion-Dollar Startup — and Its Ripple Effects

Sherwin thinks people aren't pricing in the second and third order effects of extreme individual leverage. The obvious first-order effect is the “one-person billion-dollar startup.” But the real implications are:

Second order: If one person can build a billion-dollar company, it's also way easier for anyone to build any startup. Sherwin predicts a “huge startup boom and SMB-style boom” — anyone can build vertical software. We might enter a “golden age of B2B SaaS.”

Third order: For every one-person billion-dollar startup, there might be “a hundred other small startups building bespoke software” to support it. The $10M-$50M individual business becomes life-changing for founders. But this could shrink venture-scale returns — the VC ecosystem changes when most companies are smaller.

03Theme 3: Why Most Enterprise AI Deployments Have Negative ROI

Sherwin observes that many companies are deploying AI top-down without bottom-up adoption. The anti-pattern: CEO mandates “AI-first,” ties it to performance reviews, but employees have no idea what to do with the tools and no one to learn from.

The companies where it works have both top-down buy-in AND bottom-up evangelism. They staff internal “tiger teams” — often not software engineers, but “software-engineering-adjacent” people like an operations lead who's an “Excel wizard” — who explore capabilities, build best practices, run hackathons, and create internal excitement.

04Theme 4: The Models Will Eat Your Scaffolding for Breakfast

Sherwin cites a quote from Nicholas, founder of Finool: “The models will eat your scaffolding for breakfast.” In 2022-2023, everyone built elaborate scaffolding around weak models — vector stores, agent frameworks, complex embedding pipelines. As models improved, much of that scaffolding became unnecessary. File-based skills, agents.md files, and simple search replaced complex RAG pipelines.

Even current scaffolding (skills files, context management) may eventually be eaten. Sherwin's advice: “Make sure you're building for where the models are going and not where they are today.” The best startups build products for a capability level that's ~80% of the way there today — then when the model improves, the product suddenly clicks.

He connects this to the “bitter lesson” — the less you over-complicate AI systems, the more they scale. Even OpenAI's own API team has been “guilty of this” — taking architectural turns they shouldn't have.

05Theme 5: Management in the AI Era

Sherwin's management philosophy centers on the “surgeon” analogy from The Mythical Man-Month: your job as a manager is to make every engineer feel like a surgeon — empowered, with an army of support around them (even though it's really just you).

In the AI world, this means spending 50%+ of your time with top performers. Codex supercharges high-agency engineers — the productivity spread on teams is widening. Managers should focus on unblocking top talent, especially from organizational/process blockers.

He also predicts managers will manage much larger teams. Tools like ChatGPT with organizational knowledge (hooked up to GitHub, Notion, docs) can help managers understand context better, run performance reviews, and anticipate blockers.

06Theme 6: Business Process Automation — The Underrated Opportunity

Sherwin is “extremely bullish” on business process automation, which he thinks is underrated because Silicon Valley focuses on open-ended knowledge work. Most of the economy runs on repeatable business processes — support operations, utility companies, standard operating procedures. AI can automate these with high determinism, and the impact could be even bigger than software engineering.

07Theme 7: Where the Platform Is Heading

Two major trends Sherwin is excited about over the next 12-18 months:

1. Longer coherent tasks: The “meter benchmark” shows models can already handle multi-hour software engineering tasks 50% of the time. In 12-18 months, models may handle multi-hour tasks reliably, and eventually “six-hour-long tasks where you dispatch it and have it do things on its own for a while.”

2. Audio and multimodal models: Audio is “hugely underrated” in enterprise and business settings. A lot of the world's business happens via audio, and native speech-to-speech models are improving rapidly.

On the platform stack: Responses API (lowest-level primitive for long-running agents), Agents SDK (scaffolding for multi-agent orchestration), and Agent Kit (UI components for deploying agents).