Maybe we're keeping humans in the wrong part of the AI coding loop
RoyPartha04 · reddit · 2026-11-01
A Reddit long-read argues the popular "AI proposes, human approves" loop is misplaced: let agents run autonomously and spend the saved hours on verification and testing instead.
Key points:
- Plans matter more than ever: clear goals, constraints and conventions reduce AI hallucination and make its unfamiliar code easier to understand.
- Let AI cook: an agent that does a week of work in hours shouldn't be paused for review at every step — invest that time in breaking and evaluating what it built.
- Human review isn't automatically better: LLMs draw on patterns (distributed systems, security, concurrency) most individual developers have never encountered.
- The proposed workflow: define goals in plain English → agent proposes architecture → enforce coding conventions → require unit/integration/e2e tests, static analysis, linting, security checks → agent implements and fixes → human verifies behavior, challenges assumptions, tests edge cases.
Security, data integrity and architectural decisions still deserve human scrutiny. The author asks: review every change, or grant agents autonomy and focus on rigorous testing?
More from coding & agent
- FrankenSim dev shares a prompt trick: swap physics terms to boost simulation realism — doodlestein · 2026-11-01
- claude -p burns ~1.8× more subscription quota than Claude Code TUI, Magpie switches — dotey · 2026-11-01
- LangChain's Harrison Chase: new context engineering challenges as agents go proactive — hwchase17 · 2026-11-01
- guizang-pages-skill: turn Markdown into polished single-file HTML pages via AI agents — op7418 · 2026-11-01
- One Pipecat pipeline, six voices: classifier-fronted handover between AI characters, one-shotted by Opus — alexcovo_eth · 2026-11-01
- Build a 5-bot Grok content crew: drafts, video, replies and calendar on autopilot — nima_owji · 2026-11-01