AI is moving from chat to agents
Cloudflare expanded its Agent Cloud with Dynamic Workers and model access to support AI systems that run tasks and call APIs rather than just answer prompts. (itbrief.co.nz) OpenAI’s notes on GPTs and model pickers show product-level moves to let tools pick the right model for task execution rather than expose raw models to end users. (help.openai.com)
An artificial intelligence agent is a chatbot with keys: it can call software, write code, and finish a task after the user stops typing. Cloudflare and OpenAI are both shipping products around that idea in April 2026. (markets.ft.com) (help.openai.com) Cloudflare said on April 13 that it is expanding Agent Cloud with infrastructure for “build, deploy, and scale” work, including Dynamic Workers for running code in isolated sandboxes across its network. The company said the goal is to move agents from laptop demos to “production-grade workloads.” (markets.ft.com) (developers.cloudflare.com) A sandbox is a locked room for code: the program can do its job without getting broad access to the rest of the system. Cloudflare said Dynamic Workers start on demand, can block or intercept network access, and can be configured with specific bindings and storage instead of a full container. (blog.cloudflare.com) (developers.cloudflare.com) Cloudflare is pitching that design as cheaper and faster than putting every agent inside its own always-on container. In its documentation and blog, the company said container setups can take hundreds of milliseconds to boot, use hundreds of megabytes of memory, and often need to stay warm to avoid delays. (blog.cloudflare.com) The company is also framing agents as software that takes multi-step actions, not just software that answers prompts. Its April 13 announcement said the first wave of artificial intelligence centered on chatbots, while the next wave includes coding agents and tools that read context, reason, and act. (markets.ft.com) OpenAI’s product notes point in the same direction, but from the user interface side. ChatGPT release notes on March 27 said Box, Notion, Linear, and Dropbox apps gained new actions, including write capabilities where supported, so ChatGPT can do more than retrieve information from connected services. (help.openai.com) OpenAI’s older GPTs launch in November 2023 already described the pattern: custom versions of ChatGPT can combine instructions, files, and external actions so they can search, analyze data, or interact with third-party services. That turns the model into a front end for tools, not just a text generator. (openai.com) The model itself is also becoming less visible to end users. OpenAI’s model-selection guide tells developers to balance accuracy, latency, and cost, and its ChatGPT Enterprise and Edu documentation says the “Auto” option in the model picker switches between model types automatically. (developers.openai.com) (help.openai.com) OpenAI has also widened model choices for custom GPT builders inside managed workspaces. Its Enterprise and Edu release notes on April 9 said builders can choose from the full set of ChatGPT models for Custom GPTs, which lets one tool be tuned for different workflows without asking every end user to understand the model lineup. (help.openai.com) That is the common shift underneath both announcements: the visible product is still a chat box, but the real work is moving behind it into software that picks a model, calls an application programming interface, runs code in a sandbox, and returns a finished result. (markets.ft.com) (developers.openai.com)
