Thiago Cardoso builds the AI agents inside iFood Pago, the fintech arm of iFood and one of Brazil's largest technology companies.

In this episode he breaks down how iFood turns a messaging app into a bank: an agentic "financial manager" that lets a restaurant owner pay employees, settle invoices, and move money over PIX by sending a WhatsApp message or snapping a photo. We get into the engineering reality behind it, from context engineering and evals to why he thinks of agents as a graph that grows rather than a system you design up front.

This one is for anyone building AI agents in production, especially in fintech, payments, or any domain where trust and money raise the stakes.

What we cover:

- Banking over WhatsApp: how ChatBank acts as an AI bank manager for small restaurants

- Pay by photo: snapping an invoice and letting the agent handle the PIX transfer

- The control-to-magic spectrum: choosing how much autonomy a product should have

- Personalization at scale: talking to every user in their own context

- Trust and money: why an agent's mistake with $2 is a very different problem than a wrong pizza

- Agentic UI and MCP UI: rendering custom components instead of plain chat

- Context engineering and iFood's Large Commerce Model (LCM)

- Diminishing returns: how much context actually helps the task

- From one agent to a graph: breaking a complex agent into subagents

- Why architecture should "grow" through evals, not be planned in advance

- "Complexity should never be planned": the philosophy behind effective agent teams

- Self-developing agents, OpenClaw, and the expanding security and prompt-injection attack surface

- When software becomes free: what changes when code costs almost nothing

- The assembly-to-Fortran leap: AI as the next jump up the abstraction ladder

Links and resources:

- iFood Pago: https://www.instagram.com/ifood.pago/

- iFood: https://www.ifood.com.br/

- Thiago Cardoso: add guest LinkedIn / X here

- MLOps Community: https://mlops.community/

Chapters:

00:00 Agents on a spectrum: control vs magic

00:18 What iFood Pago actually is

01:10 Lending to small restaurants other banks won't

02:52 ChatBank: an AI bank manager on WhatsApp

05:00 Paying a bill by snapping a photo with PIX

06:23 What it looks like behind the scenes

08:00 Beneficios: agents for corporate meal vouchers

08:52 Scaling personalization with agents

10:42 The control-to-magic spectrum in product design

13:13 Trust, money, and the cost of an agent's mistake

15:04 Agentic UI components and MCP UI

19:12 Context engineering and the Large Commerce Model

24:53 Breaking one agent into a graph of subagents

26:39 How agent architecture grows through evals

29:13 Why complexity should never be planned

30:25 OpenClaw, self-developing agents, and security risk

36:45 What happens when software becomes free

41:52 The assembly-to-Fortran abstraction leap

#AIAgents #Fintech #iFood

The Control-vs-Magic Spectrum Building Agents

Agentic Conversations (formally mlops.community)

The Control-vs-Magic Spectrum Building Agents