ChatGPT voice tasks and the case for agent guardrails
ChatGPT is adding voice-triggered work tasks for Plus and Pro subscribers, while Meta is planning more ways for its Muse agent to shop and handle work. Two new developer toolkits put the focus on permissions and human review, and ChatGPT Ads is expanding across Southeast Asia and Taiwan. The common question: what can an AI assistant do on your behalf, and who checks its work?
The Big Story
TechCrunch AI
ChatGPT mobile app gets voice-based agentic features
OpenAI is bringing voice-triggered tasks to the ChatGPT mobile app’s Work tab for Plus and Pro subscribers. They’ll be able to dictate requests to draft emails or documents, summarize Slack messages, create presentations and websites, and use a cloud browser. OpenAI also says users can move a conversation between phone and desktop, while Free and Go users get access to plugins and connected apps rather than the same Work features.
If you already pay for Plus or Pro, try a low-risk voice workflow on mobile, but review any draft and check connected-app access before using customer or financial information.
For Everyone
TechCrunch AI
Everything new coming to Meta’s AI agent Muse
Meta announced new capabilities for its Muse AI agent at Connect, including planned Mac app control, an email address for delegated threads, and voice access through its smart glasses in the coming months. Muse will also connect to shopping and payment services including Shopify, Shop Pay, PayPal, and Stripe; Meta says developers can build connectors for other businesses. Zuckerberg said Muse would be free for a large number of tokens, with Meta eventually taking a small fee on transactions, but the article gives no launch date for the Mac or email features.
If your business uses Shopify, watch how Muse presents products and handles purchases; wait for launch details and permissions controls before letting it operate work apps or email.
For Builders
GitHub Trending
harness-sdk: Build an agent harness and control it end-to-end. Open-source SDK for production AI agents in Python & TypeScript - any model, any cloud.
Strands Agents offers an open-source toolkit for building AI agents in Python and TypeScript, with a ready-made harness for teams that don't want to assemble the agent loop themselves. It runs inside the developer's application without a hosted control plane and supports model providers including Amazon Bedrock, Anthropic, OpenAI, and Gemini. The toolkit includes execution limits, memory, tool connections, guardrails, tracing, and evaluation; developers can use the preconfigured harness or customize the lower-level SDK.
This is a developer toolkit, not an AI app a small business can switch on today. If you're commissioning a custom agent, ask your developer how they'll cap usage, inspect agent actions, and protect customer data before choosing a framework.
GitHub Trending
agent-native: A framework for building agentic apps
BuilderIO released Agent-Native, an MIT-licensed TypeScript framework for building AI agents with a user interface where people can inspect and approve work. Developers define an action once for both the UI and the agent, sharing validation, permissions, and data rather than having the agent click through screens. The project includes authentication, scheduled automations, multi-agent teams, and PostgreSQL support, with PGlite for local development.
This is a developer framework, not a ready-made business assistant. If you’re commissioning a custom agent app, ask whether its actions have shared permission checks and a clear human review interface.
GitHub Trending
superpowers: An agentic skills framework & software development methodology that works.
Superpowers is an open-source set of skills that gives AI coding agents a structured development workflow. It asks for a design before coding, turns the approved design into small implementation tasks, and uses test-first development and code review as the agent works through them. The project supports several coding tools, including Claude Code, Codex, Cursor, GitHub Copilot CLI, and Pi, but requires a separate installation for each.
This is for businesses already using AI to write software, not a general-purpose office AI tool. If that's your team, test it on a small project and check whether the extra design and review steps catch mistakes without slowing delivery too much.
For Business
OpenAI News
ChatGPT Ads expands to Southeast Asia and Taiwan
OpenAI is rolling out ChatGPT Ads in Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam, and Taiwan, bringing the ad platform to more than 60 countries. Eligible businesses can use self-service Ads Manager or work through OpenAI and agency partners. Ads appear only for Free and Go users; OpenAI says paid Plus, Pro, and Enterprise accounts remain ad-free, conversations stay private from advertisers, and ads don't affect answers.
Businesses selling in these markets can check Ads Manager eligibility, but should test a small campaign and verify measurement options before shifting spend from established channels.