# Codex Goes Mobile and Practical Agent Loops Get Sharper *By Coding Agents Alpha Tracker ā€¢ May 15, 2026* OpenAI's Codex mobile preview was the clearest workflow shift today: coding agents are becoming remote operators you can steer from anywhere. Inside: Riley Brown's deploy-on-every-change setup, Thibault Sottiaux's recurring Codex workflows, and the key releases and harness signals worth tracking. ## šŸ”„ TOP SIGNAL - **Codex just moved from "terminal tool" to "remote operator."** OpenAI's preview puts Codex inside the ChatGPT mobile app so you can start work, review outputs, steer execution, and approve next steps from iOS/Android while the agent keeps running on your laptop, Mac mini, or devbox [^1][^2]. Greg Brockman calls it a "huge step forward for universal usage of agents," and Riley Brown's day-one demo makes that concrete with voice prompting, long-running task notifications, and deploy-on-every-change app iteration from the phone [^3][^4]. ## āš” TRY THIS - **Set up Codex mobile as a real remote console.** Riley Brown's exact sequence: update Codex desktop and ChatGPT iOS, restart Codex, keep both on the same Wi-Fi, connect from the mobile prompt, and authorize the same account so chats sync [^4]. Then switch to **Chats first** for one-off/non-coding tasks, keep **Projects** for coding, use **voice mode** for hands-free prompting, and leave notifications on because some agents run **10-30 minutes** on longer jobs [^4]. Plugins are available via `@`; skills do not show up in the mobile picker yet, but Brown says natural-language requests still hit desktop-side skills, and if you're comfortable with it he recommends **full access** permissions instead of approving every action [^4]. - **Create a deploy-on-every-change skill for phone-based app iteration.** Brown's setup: enable the Vercel plugin in Codex desktop, then prompt: `make a skill called YOLO Mode ... every single change is deployed to prod ... the public link is sent` [^4]. After that, add "please YOLO it" to prompts; he shows it creating a landing page, returning a public link, and reusing the same link after a dark-mode revision request [^4]. He also demos `create a full notes app on YOLO mode. Mobile optimized. Use Supabase for DB... like Trello` and gets a deployed app with auth and persisted data [^4]. - **Schedule a daily "chief of staff" agent.** OpenAI Codex lead Thibault Sottiaux says he hands off **100+** tasks/day and runs a recurring automation: have Codex go through **Gmail, Notion, and calendar**, summarize the day, flag what is at risk, and schedule it for **9am daily** so the report lands in the inbox [^5]. This is a good pattern if your real bottleneck is launch coordination, on-call visibility, or project drift rather than writing the next function. - **Prompt like a manager; review like a database engineer.** Sottiaux says the biggest lift comes from defining what "good" and "solved" look like, including exact output structure [^5]. Mike Krieger says he now hashes out the spec with Claude *before* it writes code so the model converges on a clear North Star [^6]. But Simon Eskildsen still manually reviews every line touching TurboPuffer's database, and Peter Steinberger's looped `codex /review` skill explicitly stops short of architecture decisions ā€” good reminder to use agent loops for local cleanup, not irreversible system design [^7][^8]. ## šŸ“” WHAT SHIPPED - **Codex in ChatGPT mobile app (preview).** Start work, review outputs, steer execution, and approve next steps from mobile while Codex runs on laptop/Mac mini/devbox; Greg Brockman called it a "huge step forward for universal usage of agents," and Romain Huet says it's live on iOS and Android [^1][^3][^2]. [OpenAI post](https://x.com/openai/status/2055016850849993072) [^2]. - **OpenClaw v2026.5.12.** OpenAI setup now defaults to **Codex** login; runtime fallbacks and stalled-stream recovery were added; Telegram polling survives stalls; installs/startup got leaner/faster [^9]. Steipete says the team has been pushing performance, reliability, security, stability, new **crabbox** automation flows, and automated video QA [^10]. [Release notes](https://github.com/openclaw/openclaw/releases/tag/v2026.5.12) [^9]. - **CodexBar 0.26.0.** New integrations for **Kiro, Antigravity, OpenRouter, Kimi**; calmer menus + keyboard nav; better **Codex/Claude** limits and cost scoping; macOS asset and CLI/Homebrew fixes [^11]. [Release](https://github.com/steipete/CodexBar/releases/tag/v0.26.0) [^11]. - **mcporter 0.11.0.** Steipete says he now uses it mainly as a more stable browser-automation CLI and for agents testing MCPs without restarts; he expects it to matter less as **code mode** spreads across harnesses [^12]. [Release](https://github.com/openclaw/mcporter/releases/tag/v0.11.0) [^12]. - **Harness quality signal from Theo.** His current ranking is rough on Claude Code: he calls it the worst harness, says **OpenCode** has better UX, multi-model support, and cheaper/faster tool-call pruning, says **Cursor** performs better with Opus, and notes that most of his own T3 Code usage is with **Codex 5.5 low/fast** anyway [^13]. - **xAI Grok Build (early beta).** New CLI for coding, app building, and workflow automation for SuperGrok Heavy subscribers; xAI says the beta is meant to improve from user feedback, and Theo's immediate take was "fast and flicker-free" [^14][^15]. [Try it](http://x.ai/cli) [^14]. - **LangChain Deep Agents 0.6.** New **harness profiles** for open models, **code interpreter** inside the loop, streaming typed projections, **DeltaChannel** checkpoints, and **ContextHubBackend** for skills/policies/memories [^16][^17]. [Blog](https://www.langchain.com/blog/deep-agents-0-6) [^16]. ## šŸŽ¬ GO DEEPER - **08:35-09:17 ā€” Riley Brown: YOLO Mode setup.** Best short clip if you want phone-native vibe coding today: Brown shows the exact Vercel-plugin + custom-skill prompt that makes every change auto-deploy to a public URL [^4]. [![Codex Mobile Released and It's Insane](https://img.youtube.com/vi/4OiLljSo-hw/hqdefault.jpg)](https://youtube.com/watch?v=4OiLljSo-hw&t=514) *Codex Mobile Released and It's Insane (8:34)* - **17:57-19:05 ā€” Thibault Sottiaux: Codex as daily chief of staff.** Good clip for anyone thinking beyond codegen. He walks through handing Codex recurring coordination work across Gmail, Notion, and calendar, then scheduling the summary for 9am daily [^5]. - **27:36-28:37 ā€” Thibault Sottiaux: define "done" precisely.** Short, reusable prompting lesson: make the output shape explicit and help Codex evaluate its own success instead of giving a fuzzy objective [^5]. - **46:40-47:53 ā€” Mike Krieger: spec before code.** Worth watching if your agents keep producing something technically correct but strategically off. Krieger's fix is to collaborate on the spec first, then let Claude implement against a crisp North Star [^6]. - **Repo to study ā€” [steipete's codex-review skill](https://github.com/steipete/agent-scripts/blob/main/skills/codex-review/SKILL.md).** Tiny repo, important pattern: iterative review loops are cheap now, but the author explicitly warns they do not replace architecture judgment [^8]. - **Release notes to study ā€” [OpenClaw v2026.5.12](https://github.com/openclaw/openclaw/releases/tag/v2026.5.12).** Read this if you care about real-world agent reliability; the changelog is mostly about recovery paths, defaults, and startup friction, which is where production agent systems actually leak time [^9]. *Editorial take: the durable edge right now is not "more autonomous" by itself ā€” it's better remote control, tighter review loops, and much clearer task definitions.* [^1][^6][^5][^8] --- ### Sources [^1]: [š• post by @OpenAI](https://x.com/OpenAI/status/2055016850849993072) [^2]: [š• post by @romainhuet](https://x.com/romainhuet/status/2055037101041909846) [^3]: [š• post by @gdb](https://x.com/gdb/status/2055034165968384099) [^4]: [Codex Mobile Released and It's Insane](https://www.youtube.com/watch?v=4OiLljSo-hw) [^5]: [Codex for Everyday Work: AI Agents Beyond Coding](https://www.youtube.com/watch?v=DLP9CagE3dU) [^6]: [What Anthropic is building after Claude Code](https://www.youtube.com/watch?v=ktKS2MIiAmY) [^7]: [Simon Eskildsen on scaling Shopify, building turbopuffer, and the future of databases](https://www.youtube.com/watch?v=bWyOyyrVIXk) [^8]: [š• post by @steipete](https://x.com/steipete/status/2054850632067019173) [^9]: [š• post by @openclaw](https://x.com/openclaw/status/2055013211473154309) [^10]: [š• post by @steipete](https://x.com/steipete/status/2055026017291370701) [^11]: [š• post by @steipete](https://x.com/steipete/status/2055163690790334865) [^12]: [š• post by @steipete](https://x.com/steipete/status/2054986075232199038) [^13]: [Iā€™m done.](https://www.youtube.com/watch?v=131yAOjxHHQ) [^14]: [š• post by @xai](https://x.com/xai/status/2054993285152989373) [^15]: [š• post by @theo](https://x.com/theo/status/2055081594537537941) [^16]: [š• post by @LangChain_OSS](https://x.com/LangChain_OSS/status/2054641656222388700) [^17]: [š• post by @LangChain](https://x.com/LangChain/status/2055018100870066414)