We’re in the middle of another leap in abstraction.
Like compilers, cloud, and containers before it, AI coding agents arrived with hype, fear, and broken assumptions. We gave the monkeys GPUs. Sometimes they output Shakespeare. Other times, they confidently ship code that compiles, passes tests, and still does the wrong thing.
The problem is simple: intent gets lost between what we mean, what we ask for, and what actually runs.
This talk delivers a practical model for software development with AI coding agents built on three equally essential ideas:
The Chasm: the divide between human intent and what is actually expressed to an AI coding agent.
The Context: the shared, explicit, and reusable knowledge an AI coding agent operates within. APIs, conventions, constraints, and domain rules replace guessing.
The Chain: the Intent Integrity Chain. A structured flow of prompt → spec → test → code, at each stage produces a verifiable artifact and is validated externally and grounded in a shared context at every stage.
Together, these form a system where intent survives implementation. Natural language becomes specifications. Specifications become tests. Tests become code. Every step is grounded in a shared context instead of assumptions and is never validated by the same model. This approach is informed by recurring failure patterns observed in real AI agents development workflows: systems passed tests, shipped successfully, yet still failed to meet intent.
Like compilers, cloud, and containers before it, AI coding agents arrived with hype, fear, and broken assumptions. We gave the monkeys GPUs. Sometimes they output Shakespeare. Other times, they confidently ship code that compiles, passes tests, and still does the wrong thing.
The problem is simple: intent gets lost between what we mean, what we ask for, and what actually runs.
This talk delivers a practical model for software development with AI coding agents built on three equally essential ideas:
The Chasm: the divide between human intent and what is actually expressed to an AI coding agent.
The Context: the shared, explicit, and reusable knowledge an AI coding agent operates within. APIs, conventions, constraints, and domain rules replace guessing.
The Chain: the Intent Integrity Chain. A structured flow of prompt → spec → test → code, at each stage produces a verifiable artifact and is validated externally and grounded in a shared context at every stage.
Together, these form a system where intent survives implementation. Natural language becomes specifications. Specifications become tests. Tests become code. Every step is grounded in a shared context instead of assumptions and is never validated by the same model. This approach is informed by recurring failure patterns observed in real AI agents development workflows: systems passed tests, shipped successfully, yet still failed to meet intent.
Baruch Sadogursky
Tuxcare
Baruch Sadogursky (@jbaruch) did Java before it had generics, DevOps before there was Docker, and DevRel before it had a name. He started DevRel at JFrog when it was ten people and took it all the way to a successful $6B IPO by helping engineers solve problems. Now, Baruch keeps helping engineers solve problems, but also helps companies help engineers solve problems. He is a co-author of the "Liquid Software" and "DevOps Tools for Java Developers" books, Java Champion and CNCF Ambassador alumni, serves on multiple conference program committees, and regularly speaks at numerous most prestigious industry conferences, such as DevNexus, DevOpsDays, Voxxed Days, Devoxx, DevRelCon, Kubecon and QCon. Today, he's taking care of developers at TuxCare.