CodeScene's Agentic Development Journey

Refactoring Street Fighter with AI | Adam Tornhill

Topic 5 of 6

Published Oct 2, 2026 · Clip 32:39–48:30 (15:51)

Adam recounts CodeScene's journey to fully agentic development: one team started a year ago building their code MCP in Clojure, achieving 2–3x lead time speedups. The tipping point was rewriting the Python-based MCP to Rust using an agent over three hours, possible thanks to strong end-to-end tests and healthy code. The team leans on deterministic safeguards: code health checks, strict code coverage gates, architectural and domain-specific dependency checks. Processes vary by task uncertainty: bugs are handled with a failing test case as evidence, no code review; open-ended tasks start from a spec, iterate on AI-generated end-to-end tests (where Adam spends 90% of review effort), commit tests before implementation, then drive feature by feature. Adam emphasizes keeping thinking human so the team builds knowledge, which is virtually impossible to enforce at scale except with a small, super-senior team. He also agrees static typing helps LLMs by constraining the solution space.