further reading
Resources
Talks, papers and articles on building software with AI agents, each picked because it helps with a lesson.
Videos
Talks to watch. Each one opens on YouTube.
Andrej Karpathy: Software Is Changing (Again)
Y Combinator, 2025youtube.com
How language models change what software is and how it gets written, and why people should stay in control.
Does AI Actually Boost Developer Productivity? (100k Devs Study) - Yegor Denisov-Blanch, Stanford
AI Engineer, 2025youtube.com
A Stanford researcher presents productivity findings from a large dataset of developers working with and without AI.
Mastering Claude Code in 30 minutes
Anthropic, 2025youtube.com
A practical walkthrough of working with a command-line coding agent, including how to give it project context.
No Vibes Allowed: Solving Hard Problems in Complex Codebases – Dex Horthy, HumanLayer
AI Engineer, 2025youtube.com
A research, plan, then implement way of using coding agents on large existing codebases, with care for context.
TDD, AI agents and coding with Kent Beck
The Pragmatic Engineer, 2025youtube.com
The creator of test-driven development on how he works with AI coding agents, and where tests fit in.
The New Code — Sean Grove, OpenAI
AI Engineer, 2025youtube.com
Argues that written specifications, not code, are becoming the main thing engineers produce and maintain.
Papers
Research to read. Each one opens on the publisher's page, free to read.
Emerson Murphy-Hill, Jenna Butler, Alexandra Savelieva, 2026arxiv.org
How coding agents spread among engineers, and roughly 24% more merged pull requests, with caveats about that measure.
Lesson: Cost and the numbers to watch
AI Writes Faster Than Humans Can Review: A Longitudinal Study of an Enterprise 2x Mandate
Hao He et al., 2026arxiv.org
When a company pushed AI to double output, the load per code reviewer doubled and automated review overtook human review.
Lesson: What changes for the team
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
Joel Becker et al., 2025arxiv.org
In a randomised trial, AI tools made experienced developers 19% slower, though afterwards they estimated 20% faster.
Mutation-Guided LLM-based Test Generation at Meta
Christopher Foster et al., 2025arxiv.org
Plants targeted faults and has a model write tests that catch them; engineers accepted 73% of those tests.
Automated Unit Test Improvement using Large Language Models at Meta
Nadia Alshahwan et al., 2024arxiv.org
A generated test is kept only if it passes filters showing it measurably improves the existing tests.
SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
John Yang et al., 2024arxiv.org
Shows that coding agents do better with tools and interfaces built for them, and measures by how much.
Joseph Spracklen et al., 2024arxiv.org
Code models often name packages that do not exist, which attackers can exploit; measured across 16 models.
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
Carlos E. Jimenez et al., 2023arxiv.org
How 2,294 real GitHub issues from 12 Python projects became the common test of whether a model can fix real code.
Do Users Write More Insecure Code with AI Assistants?
Neil Perry et al., 2022arxiv.org
In a user study, people with an AI assistant wrote significantly less secure code than those without one.
Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions
Hammond Pearce et al., 2021arxiv.org
Measures how often an AI code assistant suggests code with known kinds of security weakness.
Articles
Guides from teams that build this way. Each one opens on its own site.
Anthropic, 2026claude.com
A stage-by-stage guide to planning, designing, building, testing, deploying and maintaining software with an AI agent.
Lesson: It is a loop, not a line: plan, build, ship, watch, learn