A start-to-finish pass through creating an agent, with no code written at any point. Vinit Agrawal signs in, chooses to create an agent from a Knowledge Base, points it at a public website, and watches the crawler pull the pages in. The configure view lists every page that got indexed and the text taken from each one, so you can read the content exactly as the agent will read it.

From there he opens up the parts of the build that decide quality. The conversation is broken into Gambits, with a static welcome message first and an agent Gambit holding the prompt, the model, the connected tools, and settings like whether to use conversation history. The prompt is treated as the crucial piece, and rather than hand-writing one he uses the generator: a short plain-English instruction about what the agent should do becomes a detailed prompt plus a welcome message, both editable before you apply them. He then asks the preview agent a real question and shows the knowledge retrieval tool being called, along with the settings behind it for how many chunks get retrieved and how many are passed to the model. The second half switches from answering to acting. A Google Sheets tool is connected, its sub-tools shown individually so unnecessary ones can be switched off, and an enrollment request captures a name, email, phone number, and interests into a live sheet while the tool calls stay visible in the thread.

The last piece is jump logic, and it is the one worth watching twice. Instead of a coded rule, the branch out of the agent Gambit carries a plain-English success condition, in this case that the user has asked to speak to a person. He publishes, types the request with a typo in it, and the agent finishes its answer and then hands control to the handover Gambit. Watch this if you want to see what a working agent is actually made of, end to end, before you open your own account.