Most important take away

We crossed a critical inflection point in November 2025 when GPT 5.1 and Claude Opus 4.5 became reliable enough that coding agents consistently produce working code, fundamentally changing how software is built. The biggest shift is not just speed but rethinking every process around software development now that the expensive part (writing code) is nearly free, while the hard parts (deciding what to build, verifying quality, and security) become the new bottlenecks.

Chapter Summaries

The November Inflection Point

Throughout 2025, Anthropic and OpenAI focused training efforts on code and reasoning. In November, GPT 5.1 and Claude Opus 4.5 crossed a threshold where coding agents went from “mostly works” to “almost always does what you asked.” Engineers returning from holidays realized the technology had gotten dramatically better, sparking widespread adoption.

Vibe Coding vs. Agentic Engineering

Simon distinguishes between vibe coding (not looking at code, great for personal prototypes) and agentic engineering (professionals using coding agents to build production software with review and quality standards). He argues the term “vibe coding” loses meaning if it covers everything AI-assisted.

The Dark Factory Pattern

Companies like StrongDM are experimenting with “nobody reads the code” policies, relying instead on swarms of AI-simulated QA testers running 24/7. They built simulated versions of Slack, JIRA, and other tools to test their security software at scale, spending roughly $10,000/day on tokens but achieving robust continuous testing.

Where Human Value Remains

The bottleneck has shifted from code production to deciding what to build and proving ideas work. Simon prototypes three different approaches for every feature because code is so cheap. Usability testing with real humans remains essential since AI cannot credibly simulate actual user behavior.

The Impact on Engineering Careers

Experienced engineers benefit most because AI amplifies 25 years of skill. New engineers benefit from faster onboarding. Mid-career engineers are in the most precarious position according to a ThoughtWorks study, lacking the deep expertise to amplify and already past the beginner onboarding boost.

Mental Exhaustion and the Productivity Paradox

Despite AI making work faster, the most AI-engaged engineers report being mentally wiped out by mid-morning. Running four parallel agents requires intense cognitive overhead. Simon hopes this exhaustion is a novelty effect but acknowledges the gambling-like addictiveness of always wanting agents running.

Agentic Engineering Patterns

• Code is cheap now: The thing that used to take the longest (writing code) now takes minutes, so the focus shifts to making that code good.
• Hoard things you know how to do: Maintain repositories of small tools, research projects, and experiments as reusable context for future AI-assisted work. Simon keeps 193+ HTML/JS tools and 75+ research projects on GitHub.
• Red/green TDD: Tell agents to use “red/green TDD” as shorthand for writing tests first, watching them fail, implementing, then watching them pass. This five-word prompt dramatically improves agent output quality.
• Start with good templates: A thin project skeleton with a single test and preferred formatting style is enough for agents to pick up and follow consistently.

AI Tools and Stack

Simon primarily uses Claude Code (especially the web/phone version for safety and convenience) and has been leaning into GPT 5.4 recently. He uses AI for research/search more than Google directly. He avoids memory features to maintain objectivity as a researcher.

The Lethal Trifecta and Challenger Disaster of AI

Prompt injection remains unsolved. The “lethal trifecta” occurs when an agent has: (1) access to private information, (2) exposure to malicious instructions, and (3) a way to exfiltrate data. Simon predicts a “Challenger disaster” moment where the normalization of deviance in AI security practices leads to a major, headline-grabbing breach. The CaMeL paper from Google DeepMind offers a promising architectural approach but no robust implementations exist yet.

OpenClaw and Personal AI Assistants

OpenClaw went from first line of code (November 25) to a Super Bowl ad in 3.5 months, proving massive demand for personal AI assistants despite serious security concerns. Simon sees building a secure version of OpenClaw as the biggest current opportunity in AI.

Simon’s Current Work

He continues building open-source data journalism tools (with a goal of contributing to Pulitzer Prize-winning reporting), is writing an agentic engineering book one chapter at a time on his blog, and recently started monetizing his blog through subtle sponsorships.

Summary

Key Themes:

• The AI coding revolution hit an inflection point in November 2025 where models became reliable enough to fundamentally change professional software engineering.
• “Dark factories” represent the next frontier: production software built without humans reading code, relying instead on automated testing swarms and quality frameworks.
• Security (especially prompt injection) remains the critical unsolved problem, with a major disaster likely inevitable due to normalization of risky practices.
• The value of human engineers is shifting from writing code to judgment, architecture decisions, and knowing what to build.

Actionable Insights:

• Use “red/green TDD” as a standard prompt when working with coding agents to dramatically improve code quality.
• Prototype three versions of every feature since code is essentially free now, then use real usability testing to choose.
• Build and maintain personal repositories of small tools and research experiments as reusable context for AI agents.
• Start every new project with a thin template that demonstrates your preferred code style; agents will follow it.
• Run coding agents in sandboxed environments (like Claude Code for web) to limit blast radius from potential security issues.
• Cut off the exfiltration leg of the lethal trifecta when building AI-powered systems that handle private data.

Career Advice:

• Lean into AI tools rather than resisting them; the only universal skill right now is the ability to roll with rapid changes.
• Invest in your own agency (deciding what problems to tackle) since that is the one thing AI fundamentally lacks.
• Be more ambitious than you think is reasonable; projects that would have taken weeks now take hours.
• Mid-career engineers should focus on deepening expertise and building a broad backlog of techniques and experiences to amplify with AI.
• The exhaustion from intensive AI-assisted work is real; set boundaries and recognize cognitive limits to avoid burnout.

Product Strategies:

• Building a secure personal AI assistant (a safe OpenClaw) is the single biggest product opportunity in AI right now.
• AI-driven QA swarms that simulate real users 24/7 can replace traditional QA departments at scale.
• The “alpha” label matters more than ever because high-quality tests and documentation no longer signal that software has been battle-tested by humans.
• Artisanal, human-used software will be valued more as AI-generated code floods the market; proof of real usage becomes a key differentiator.