Thursday, August 13: AI starts acting on your behalf

Today’s verified AI releases share one direction: the tool is moving from giving an answer to taking a bounded action. Three launches show that shift in small-business growth, engineering automation, and real-time API work.

Delivery Hero launched an agentic assistant for shops and restaurants. Owners can use text, voice notes, or images in WhatsApp to inspect menus, spot weak-selling products, plan promotions and ads, and draft responses to customer reviews. The assistant can implement a change only after the owner approves it. Delivery Hero says it already supports more than 40,000 partners, and reports a 15% order increase for restaurants using it through Glovo. That result is a company-reported early signal, not an independent benchmark. 1

Why it matters: this is an AI agent attached to an operating channel a small business already uses. The useful question is not whether it can suggest a promotion. It is whether approval stays clear when the suggestion touches advertising, pricing, or customer communication.

PTC announced the availability of the FeatureScript MCP Server in its Onshape CAD and product-data platform. An engineer can describe a custom CAD feature in natural language, then use AI to help build, test, debug, and refine the underlying FeatureScript. The resulting automation can be reused and shared across teams, rather than regenerated from scratch each time. PTC says the connection works with coding models including Claude, ChatGPT, and Gemini, and is available through Onshape Labs. 2

Why it matters: the product is not only generating a CAD output. It is helping turn an engineer’s know-how into a repeatable tool. That makes the adoption test more concrete: can the generated feature pass engineering review and become part of a team’s normal workflow?

OpenAI previewed Ultrafast, a new API service tier for GPT-5.6 Sol powered by Cerebras. OpenAI says it can reach up to 750 output tokens per second, or up to 14 times the speed of standard processing. The target use cases are time-sensitive workflows such as incident response, complex customer support, financial research, and interactive experimentation. Access is limited to an initial group of customers while OpenAI studies where the speed changes the product itself. 3

Why it matters: faster inference is not just a nicer loading state. If the model remains capable enough, a workflow that used to wait for a batch result can become interactive. The boundary today is important: this is a limited preview, not a generally available API tier.

Taken together, today’s drops make the same bet at three different layers. Delivery Hero puts an agent inside an existing business conversation. Onshape turns engineering expertise into reusable automation. OpenAI is trying to make frontier intelligence keep pace with a live user. For a closer look, compare the approval surface, the review burden, and the actual availability—not just the headline capability.