Two working demos of the same idea, at two levels of complexity. Both solve a real people-team problem, new hire onboarding, and both follow the same architecture I use for any agent.
A grounded Q&A agent. Answers new-hire questions only from policy, cites the section it used, and routes sensitive topics to a human. One question in, one trusted answer out.
An autonomous multi-step agent. Give it a goal and it plans the steps, calls multiple tools in sequence, shows its reasoning, and assembles a final result, while escalating anything sensitive to a human.