Understanding an agent product by what it can actually do

The distinction between a chatbot and a work agent becomes clearer when you inspect its capabilities. In his August exploration, Simon Willison found that the cloud version of ChatGPT Work combined networked code execution, browser interaction, file access, website publishing, subagents and scheduled prompts.

Those parts could compose into an ongoing workflow: investigate a topic, collect structured data, build a site and later update it. The interesting unit was the connected process, not any one generated answer.

He also distinguished cloud execution from the desktop product’s local access. Similar branding did not mean identical environments.

The investigation exposed a documentation problem. Descriptions of intended use left important operational differences unclear; explicit tools and environment specifications would have been more useful. It also raised a security question: private data, untrusted web content and outbound communication occupy the same workflow. The article records what he observed then, rather than a permanent specification of a rapidly changing product.