Incidental Architecture

An airhacks.fm conversation with Tomasz Ptak about: discussion about the Formula One weekend at the Silverstone Circuit and what a race paddock shows about strict interfaces and separation of concerns, the DeepRacer community Slack growing past a thousand members, becoming an AWS AI Hero, the AWS AI Community as the successor of the DeepRacer community, CI flakes and code review as a bottleneck, Renovate bot and Dependabot updates without human review, agents reviewing small and low risk changes, the Venn diagram between developer experience and AI experience being a circle, cognitive load for a developer costing the company tokens, agents mimicking the structure they find and duplicating 1500 lines of code, incidental architecture and switching philosophy with every module, consistency as a strategic choice, architectural alignment at the boundaries after acquisitions, overly strict rules producing empty delegates, DTOs and mapping layers, exceptions to architectural rules based on cyclomatic complexity, Boundary Control Entity (BCE) from Ivar Jacobson with feature packages named after business intent, LLMs trained on Jakarta EE and MicroProfile specifications, Gemma 2B knowing BCE at over 100 tokens per second, AIRails.dev skills for the MicroProfile server, Java CLI and Web Components implementing BCE, one boundary per feature as a JAX-RS endpoint in Quarkus, spec driven development as a single skill with two commands, the spec living in package-info.java as Markdown Javadoc, jmarkdoc exporting JavaDoc as Markdown documents, EARS requirements syntax and one test per requirement, unit, integration and System Tests as the verification loop, code as a cache of the understanding of the requirements, FinOps and token cost as the reason to use what the model already knows, grep and find versus the Language Server Protocol (LSP) in Claude Code, javap decompiling jar dependencies and losing context, keeping module sources visible to the agent instead of hiding them in jars, monorepo and monobuild trade-offs, coarse grained modules instead of hundreds of microservices, microservice architecture for agents, zb as a zero dependency Java builder, quality as an amplifier in both directions, a library carrying an instruction that tells agents to delete the tests, trust and verification instead of reading every line, Kiro with one spec per business component and its requirements, inference cost as the question from the GeeCON audience, the sustainability point beyond which agents amplify maintainability

Tomasz Ptak on linkedin: tomasz-ptak