Articles from martinfowler.com

Modernizing a large retail IoT system required adapting software engineering defaults to hardware realities like intermittent connectivity, energy limits, and fleet observability.

A behavioral harness offers no reliability gain for frontier models like GPT-6 Sol but significantly improves consistency and cost-effectiveness for the cheaper GPT-6 Luna model.

Thoughtworks engineers built an airline IROps system in four days via prompting, finding their repo served as an accidental blackboard that enabled emergent agent coordination, leading to the Talwrn tool.

The author argues that in multi‑agent LLM workflows the real cost is context pollution in the orchestrator’s working memory, not token usage or parallelism, and proposes rules to protect it.

The post collects expert perspectives on AI-assisted development, emphasizing verification over generation, harness engineering, function structure, and the cultural divide between software thinking and human experience.