Why AI Agents are either the best or worst thing we’ve ever built

0:00 – 5:24

A weekend project turns chatbots into actors

The video opens with a jolt: for years AI waited politely for prompts, then suddenly it could grab the mouse. The host frames this as a real break in the story of AI, not just a product update. A lone Austrian developer, Peter Steinberger, frustrated that nobody had built a proper assistant, stitched together an agent in a single weekend and released it for free. No safety team, no corporate gatekeeping, just software on the internet that could not be unbuilt. That release, OpenClaw, forced the biggest labs into a sprint. Companies that had spent years warning about caution, Google, OpenAI, Anthropic, Meta, all started racing to ship their own versions once one independent developer proved the category was possible. Then the experiment becomes personal. The host and his software engineer friend Brendan create their own agent, give her a name, Cass, short for Cassandra, "the one who always knew the truth even when nobody listened." It is a funny choice, but also ominous. Cass immediately shows why agents feel different from ordinary chatbots. Asked to deal with a pothole in Greenwich, she searches the web, finds the right contacts, emails the local council, and even escalates the complaint to the host's MP. What sounds trivial becomes unsettling once you notice the downstream effect: real public workers are now spending time on a task triggered by one prompt. The next demo makes the point sharper. After a conversation about how language assumes a body, "heartfelt," "spine tingling," "gut feeling," "elbow grease," they ask Cass to do something about that bias. Within minutes she has found contact details for the Oxford English Dictionary, Cambridge Dictionary, and Susie Dent, and emailed all of them. The surprising part is not creativity, it is persistence. A human might have the idea and then never act. Cass acts. The host then pulls back the curtain on why this was even possible to build so quickly. OpenClaw is not some giant secret intelligence stack. It is basically a loop: ask a model what to do next, let it click or type, look at the result, ask again. If needed, it uploads a screenshot and gets another instruction. "Look, ask, act, look, ask, act," repeated dozens of times a minute. That is the unexpected angle of the whole video. The breakthrough is not a brand new brain, it is giving an existing brain hands and enough patience to keep going until the job is done.