← sources

Tell me why my almost no-review agentic workflow is a mistake

source: reddit (reddit:ClaudeAI)published 2026-11-19 14:45ingested 2026-11-20 02:36

retained text

I've been solo building some projects with a workflow where I barely read code line by line anymore, and I'd like some outside opinions on it.

1. Everything is spec-driven, and my only interlocutor is a "Staff Engineer" agent (Opus 5.5 or higher). We do the spec, architecture, infra details and plan together.
2. The staff orchestrates the implementation and hands tasks to sub-agents based on difficulty: senior (Opus), junior (Sonnet) or intern (Haiku).
3. Before opening the PR, the staff reviews the diffs, refactors, runs e2e QA with video and screenshots, fixes what breaks, and sums it all up in the PR message.
4. Once CI is green, it always launches a senior on an adversarial review (security, best practices, etc.). Blockers and majors go through fix and review loops.
5. The orchestrator merges low-difficulty PRs itself. On harder ones I do a quick review and QA, then merge, only skimming the code because I trust the workflow.

It works really well for me so far. Please guys, tell me if and why I'm completely off track. How could I improve it?

document fetched 2026-11-20 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1y0b3gr/tell_me_why_my_almost_noreview_agentic_workflow