Andrew Stellman
I take difficult technical topics and make them really easy to understand.
Through my books and courses, I’ve helped people understand coding, project management, Agile development, working with teams, leadership, and process engineering and improvement. Now I’m tackling artificial intelligence.
I’ve written six O’Reilly books with my longtime collaborator, Jennifer Greene, and our books have been translated into more than a dozen languages.
If you want to learn about what I’m working on now, here’s a good starting point: The Accidental Orchestrator. I describe how I used AI to build Octobatch and what I learned along the way.
I’m a New York-based developer, architect, author, and international speaker.
Foo Camp 2026: “It’s a Feature, Not a Bug.” Watch →
Current project
Learn how LLMs work by being each piece of one.
In this book, you learn how LLMs work by doing each part of a model’s job by hand. You choose the next word from real corpus statistics, set the weights on a neuron, and work through attention calculations. You also build a working model with AI, then run it in PyTorch and export it to Hugging Face.
Now
- Bank of AmericaSVP and Principal Engineer. I build AI agents to help developers improve the quality of their code, and I train developers across the entire Global Technology organization to use AI for development.
- O’Reilly RadarA regular series on agentic engineering and AI-driven development.
- AI-driven developmentI use open-source projects like Octobatch and the Quality Playbook to test ways of developing software with AI.
Books and Reports
Background
- Software Design and DevelopmentI’ve been writing code since the 1980s, and for the past three decades I’ve been a working software developer, continuously delivering products for different companies. Most of my work has been in financial services, including multiple Fortune 100 companies. My development work has also included designing large-scale real-time backend systems at a major investment bank.
- WritingI’ve written six top-selling O’Reilly books with my longtime collaborator, Jennifer Greene, including Head First C#, Head First PMP, and Learning Agile. Our books have been translated into more than a dozen languages, and Head First C# is now in its fifth edition. I also write for O’Reilly Radar about AI-driven development, including the Sens-AI series that became the report Critical Thinking Habits for Coding with AI.
- Teaching and trainingI’ve been teaching about software engineering for over 20 years. Since 2017, I’ve taught O’Reilly live training courses on scaling Agile and project management, coding with C#, full stack development, new technologies, and AI. I also teach AI-assisted development at Bank of America.
- SpeakingIn 2006, with one book out, Dr. Dobb’s Journal called me “one of the industry’s best and brightest,” and a few months later they flew me to India to speak in Hyderabad, Chennai, and Bangalore. You can watch my Foo Camp 2026 talk, It’s a Feature, Not a Bug.
- Managing teamsI spent 10 years managing teams at different companies, ranging from small teams of half a dozen people to large groups of over 80 people, including global teams. I’ve managed developers, architects, business analysts, requirements engineers, and project managers. A lot of my work has been about learning and teaching how to produce products of very high quality.
- ConsultingOver the past 20 years I’ve consulted for various companies, including Microsoft, MIT, and the National Bureau of Economic Research.
Articles on AI-driven development
In my O’Reilly Radar series, I write about how you can work with AI coding agents, check their code, and manage their context. Read all my Radar articles →
- The Accidental OrchestratorHow I started Octobatch to handle unreliable LLM output.
- Keep Deterministic Work DeterministicPull deterministic work out of the LLM’s hands. A pipeline went from 31% to 84% passing.
- The Toolkit PatternWrite documentation for the AI to read: AGENTS.md, CLAUDE.md, and why they work.
- AI Is Writing Our Code Faster Than We Can Verify ItApply classical quality engineering to AI-generated code: requirements first, then verification.
- AI Code Review Only Catches Half of Your BugsWrite down what your software must not do. Negative requirements catch the intent bugs structural review misses.
- Why Doesn’t Anyone Teach Developers About Context Management?Treat an agent’s context the way you treat memory: manage it deliberately.
- Your AI Agent Already Forgot Half of What You Told ItFour techniques: split discovery from documentation, handoff documents, acceptance criteria, spec documents.
- When Context CollapsesExternalize, recognize, rehydrate: save agent state to disk, detect when context has degraded, reload.
- So Long and Thanks for All the ContextKeep the agent’s working context small, with important information near the beginning and end.
- My AI Kept Pushing Me to Ship, So I Asked It WhyRecognize continuation pressure, the bias that keeps nudging you toward the next step.
- We Keep Renaming AI Coding. Here’s What I’d Call It.Why the name matters, and the case for calling it AI-driven development.
Earlier articles on coding with AI
I adapted my Radar series on the Sens-AI framework into the O’Reilly report Critical Thinking Habits for Coding with AI. The original framework article introduces the habits and links to the rest of the series.
- The Sens-AI Framework: Teaching Developers to Think with AIFive habits for coding with AI: provide context, research before prompting, frame the problem precisely, iterate deliberately, apply critical thinking to everything it produces.
- Prompt Engineering Is Requirements EngineeringWrite prompts the way you’d write a good requirement, because it’s the same job.
- Bridging the AI Learning GapWhy Head First C# teaches coding with AI, and how to learn to code without letting the AI do the learning for you.
Tools you can run today
An open-source skill that guides AI coding agents through code reviews using quality engineering techniques.
A batch orchestration tool for large-scale LLM processing, with a TUI, Monte Carlo simulation support, and AI-assisted pipeline development. I used AI to write all of its code.
A batch orchestrator for AI coding agents that runs inside your existing agent session. No server, no daemon, no admin rights.
Research
I’m researching how to apply traditional quality engineering to AI-written code. The Quality Playbook is the open-source implementation, and it has found bugs in Google’s Gson and the Linux kernel’s virtio and zram drivers.
- AI Is Writing Our Code Faster Than We Can Verify ItO’Reilly Radar, 2026. How to apply classical quality engineering to finding bugs in AI-generated code, using the Quality Playbook.
- Agent Orange exposure research and HEA-VI coauthored a paper on assessing exposure to Agent Orange and other herbicides in Vietnam, published in Environmental Health Perspectives. I developed the math for estimating exposure by integrating along spray routes, replacing an approximation that treated the herbicide as two large deposits at the endpoints. With Jennifer Greene, I built HEA-V, the software used for exposure analysis for the last 20 years.
- Science policy research with Pierre Azoulay
I built the open-source tool PublicationHarvester with the economist Pierre Azoulay to collect scientists’ publication records from PubMed for science policy research. I coauthored PublicationHarvester: An Open-Source Software Tool for Science Policy Research, published in Research Policy in 2006.
My work with Pierre also included FindRelated, which retrieves related articles from PubMed and adds them to a PublicationHarvester database.
I built Scientific Distance to compare the subject headings in pairs of scientists’ publications. The method is described in Superstar Extinction by Pierre Azoulay, Joshua Graff Zivin, and Jialan Wang.