Executor overhauls agent approvals with 4 modes, adds Gmail-style undo for any agent action
threepointone · x · 2026-11-16
Executor shipped a major overhaul of how tool approvals work, offering four connection modes:
- Automatic: all tools run on the agent's judgment, good for in-the-loop usage or highly aligned models
- Agent: destructive tool calls trigger a secondary confirmation prompt with resume, offering a bit more trust
- Browser: approval happens on an Executor page, so the agent can't auto-approve even with its own token — best for background agents with destructive tool access
- Elicitation: approval happens inside your harness, with token-level permission controls
The author also highlighted the underrated Workflows feature, which brings Gmail's 'Undo Send' pattern to any agent action just by prompting.
More from coding & agent
- Cheap-first routing: only 4 of 143 questions escalated to a frontier model, cost down 80% — xrplincoln · 2026-11-16
- NVIDIA SkillSpector on 4 popular Agent Skills repos: read the findings, not the score — viest_me · 2026-11-16
- I Shipped an App, a Site, and a CRM With Codex and Claude Code — and Learned Nothing — ThatGuyJord2020 · 2026-11-16
- Over Half of LLM Requests Go to Models Younger Than 60 Days, Vercel Gateway Data Shows — aruthesa · 2026-11-16
- Open-source Craft Anything hits 34K GitHub stars in 16 days by letting Claude Code clone any app — thisguyknowsai · 2026-11-16
- InfoQ: 'Canvas for AI Agents' proposes a 4x4 framework for measuring agent work — bibryam · 2026-11-16