Dev builds human-approval queue that blocks agents from sending AI-sounding text
paulofilip3 · reddit · 2026-10-03
A developer who uses agents for emails, client messages, and commits released reright: agents can't send anything directly—every message lands in a draft queue with full context (what it is, who reads it, why), and only an edited, human-approved version goes out; rejection stops the agent. In Claude Code it hard-blocks sends (git commits, gh, Gmail, browser typing); Codex/Gemini CLI/Copilot/Cursor are configured by the same installer but only Claude Code is well-tested. Subjects/bodies open in separate boxes with rich/preview/source modes; Claude.ai and ChatGPT work via connector (request-only). Can be paused per timeframe or folder.
More from coding & agent
- Amazon (2B) and Cloudflare (9B/27B) ship open-source decision models rivaling Jev — mark_k · 2026-10-05
- Autonomous agents only pay off where value justifies engineer escalation, dev argues — Pavel_Asparagus · 2026-10-05
- Users report agents like Codex feel noticeably slower than three months ago — koltregaskes · 2026-10-05
- Building After Effects motion graphics engines with Opus 5.5 — notiansans · 2026-10-05
- Local inference in practice: mining repos, AI chat logs and media libraries with a tiny model — natesiggard · 2026-10-05
- repOx packs 3k-file repos into LLM prompts in ~14ms, stripping 30k-token lockfile waste — Wvdy_CC · 2026-10-05