Alfonso Graziano is AI Lead at Nearform and author of AI-Native Software Engineering, an O’Reilly book about developing dependable software with AI agents. He designs systems that make agent behavior measurable, inspectable, and safe to improve in production.
His background combines JavaScript, TypeScript, Node.js, cloud infrastructure, and developer education. In 2023, he created the web-developer-roadmap, a free 15-week curriculum spanning frontend development, databases, Docker, and cloud fundamentals. By 2025, he was a Nearform technical lead working with TypeScript workflow engines; his subsequent projects include node-code-sandbox-mcp, which executes JavaScript in disposable Docker containers, and resources for AI-native engineering.
Agent evaluation must test actual behavior. Graziano builds golden datasets and evaluation scorers with domain experts, assessing answers alongside tool selection, parameters, and execution sequences. Fixed benchmarks establish baselines, expose regressions, and prevent agents from improving apparent performance by rewriting their own tests.
AutoAgent makes improvement an auditable experiment. His AutoAgent project applies ideas from Andrej Karpathy’s separate autoresearch work to agent development: a coding agent proposes a hypothesis, modifies another agent, reruns evaluations, and retains or reverses the change. Git branches, experiment reports, and persistent memory preserve successes and failures for human review.
Production failures should become regression tests. Graziano analyzes user feedback, expert annotations, execution traces, latency, and token consumption to identify recurring failure patterns. Domain specialists validate proposed fixes, and confirmed failures enter the evaluation dataset. His account of improving a client agent connects this process to practical trade-offs among accuracy, cost, and latency.
Procedural memory can preserve successful agent behavior. His experiments with procedural memory for AI agents capture effective task executions as reusable instructions that smaller models can follow, while acknowledging that those instructions can become stale or encode mistakes.
Graziano describes the surrounding discipline as harness engineering: combining specifications, contextual information, quality gates, observability, and human governance so agents can propose changes without escaping engineering accountability.
