Caio and Marina break down how to go from 'I use ChatGPT to write emails' to actually building an agent with Claude that handles repetitive tasks on its own. No fluffy theory: the right first project to pick, the mistakes that make everyone quit by week two, and the step-by-step to get an agent running this week.
In this episode
01 The hook: the difference between using AI and having an agent
- Marina opens with the audience's confusion: 'Caio, I already use ChatGPT, aren't I already doing this?' and Caio cuts in saying no, there's a clear line between a chat and an agent
- Defining an agent concretely, no jargon: a thing that gets a task, decides the steps, and runs it without you copying and pasting all day
- The episode's big turn right up front: most companies get stuck because they pick the wrong project, not because the tech is hard
- Caio gives a quick example of a client who saved hours with a dumb little email-triage agent, to anchor that this is mundane, not sci-fi
02 Picking your first project (where almost everyone messes up)
- Marina asks what she'd do as a beginner: 'I'd want to automate the most important thing in the company, right?' and Caio shows why that's a trap at the start
- Caio's three-filter test: a repetitive task, with clear rules, where getting it wrong doesn't break anything expensive
- Concrete examples: answering frequent customer questions, organizing incoming leads, summarizing a meeting and sending next steps
- Anti-example: why 'an agent that closes sales on its own' is the worst possible first project and why people get frustrated trying it
03 Hands-on: building the agent in Claude
- What Claude is and why it works well for agents: Caio explains Projects and artifacts without turning it into a tech lecture
- The practical step-by-step: give the agent a clear role, give it company context, give examples of how you want the answer, and test with real cases
- Marina keeps pushing on 'but the context, where does it come from?' and Caio shows you don't need to code, you need to write good instructions, like training a new employee
- The tools trick: connecting Claude to things it can actually access, a spreadsheet, a doc, so it doesn't just talk but actually does
04 The reality check: why your agent will mess up (and that's fine)
- Caio breaks the expectation: your first agent starts out kind of dumb and that's expected, the work is tweaking it
- How to find the mistakes fast: run ten old real cases and compare them to what you'd have done by hand
- The 'human in the loop' idea explained simply: the agent preps, you approve, and only then do you slowly let go of the reins
- Marina raises the audience's real fear, 'what if it sends something dumb to a customer?', and Caio answers with how to actually safeguard against it
05 Wrap-up: what to do this week
- Caio's practical challenge: take ONE task you did three times this week and turn it into your first agent
- The mindset mistake to avoid: wanting the perfect automation before you have the ugly one working
- Caio closes by tying it to the mission: applied AI isn't about the trendy tool, it's about getting boring work off your plate and measuring the time you got back
- Hook for the next step: what changes when you combine several small agents, leaving the audience wanting more