The Agentic Software Factory

By Hugo Bowne-Anderson (featuring Wes McKinney, Jeremiah Lowin, and Randy Olson on Show Us Your Agent Skills).

Vibe Coding vs. Agentic Engineering

“The difference between vibe coding and agentic engineering is planning, architecture, and caring about the output.” — Wes McKinney (riffing on Jesse Vincent).

When producing software at scale (e.g. 1M lines of generated code over 6 months across roughly a dozen projects at 1.3 to 1.4B tokens a day), human line-by-line review breaks down entirely. The developer’s role shifts to making structural decisions about scope, architecture, and constraints, while automated systems read the code.

Verification Stack in Practice (Wes McKinney’s Harness)

Slow Plan, Long Implement: Detailed spec interviews (e.g. via Superpowers) before implementation. Implementation runs over git worktrees for hours across dozens of tasks.
Commit Every Turn: Hard requirement in AGENTS.md/CLAUDE.md for fine-grained reviewer triggers and immutable checkpoints.
Automated Continuous Review (RoboRev): Post-commit hook firing automated code review through Codex (GPT-5.5 / high reasoning) on every commit. Findings accumulate in a per-repo ledger; code is read 4-5 times by agents before human merge consideration.
Local Tooling Ecosystem: Kata (local issue tracker), Middleman (local PR review interface), and Agents View (session search across hundreds of agent traces).

Generator-Evaluator Loops & Self-Grading Skills

Randy Olson demonstrates encoding visual and domain judgment into self-evaluating skills where the agent critiques its own output against explicit aesthetic and structural rubrics before submitting.