Humility in the Age of Agentic Coding

What It Covers

Steve Klabnik, Rust programming language contributor and author, traces his shift from AI skeptic to agentic coding practitioner. He details building the Roo programming language almost entirely with Claude, examines which software engineering beliefs hold up under AI-assisted development, and identifies the central unsolved problem of maintaining code quality at machine velocity.

Key Questions Answered

• •Agentic coding as a learnable skill: Treating AI-assisted development like Vim — a tool with a real learning curve — changes outcomes significantly. Klabnik's first Roo attempt failed partly due to his own inexperience with agentic workflows. His second attempt, with better prompting habits and tighter iteration loops, produced dramatically higher code quality and velocity.
• •Validation-first development unlocks agent performance: Klabnik built a custom test framework that connects a formal language specification directly to runnable test cases before writing compiler code. This gave Claude a concrete pass/fail signal to iterate toward, replacing manual review cycles. Agents converge on correctness faster when given automated, objective evaluation criteria rather than human spot-checks.
• •DRY and clean code conventions are human-centric heuristics: Many software engineering practices — avoiding code duplication, tab-width debates, microservice boundaries — exist to manage human cognitive limits, not machine ones. Klabnik now tolerates five identical function copies in a codebase, trusting that Claude can identify and consolidate them in seconds when it matters, rather than enforcing it upfront.
• •Non-programmers model AI uncertainty better than developers do: Klabnik's observation: software engineers are trained on determinism and treat hallucination as disqualifying. Non-technical users already expect computers to be partially wrong and fact-check outputs by default. Developers who adopt that same verification-first mindset — treating AI output as a draft, not a result — extract more practical value from the tools.
• •The unsolved problem is trust at merge velocity: Klabnik shipped roughly 100 pull requests on Christmas Day while with family, reviewing each diff in seconds rather than minutes. The productivity gain is real but the quality risk is unresolved. He frames this as the central engineering question of 2026: how to establish sufficient trust in agentic output to allow fast merging without accumulating dangerous technical debt.