In this episode of Agents at Work, Jordi Montes sits down with Yunfan, an ex-Google / ex-Meta (Llama 3) engineer now building agents at Yutori, to talk about what it actually takes to ship agents that run continuously and stay reliable in the real world.

Yunfan breaks down Scouts, Yutori’s agentic web search product that can run on a schedule (daily/weekly/hourly), browse like a human, and only notify you when something meaningfully changes without spamming your inbox.

They explore:

• What “agents” really are: a model + a loop of action → observation → next action (and why that definition matters)
• Multi-agent orchestration: orchestrator + specialized agents (travel, finance, info gathering) and how “division of labor” improves performance
• MCP vs function calling: why Yunfan thinks many tool integrations can collapse into simple functions/scripts ( “bash is all you need”) with less complexity and less context overhead 
• Agents need memory: why Yutori is evolving from a report archive to a real “file system” so agents can store progress, preferences, and reusable knowledge 
• The real problem: reliability: why 95% accuracy fails at daily cadence and the push toward 99–99.9% (plus supervision/self-healing)
• Model economy: why not all tokens are equal, and how mixing frontier models with smaller open models makes agents affordable at scale

If you’re building agents that have to run long workflows, browse the web, and deliver trustworthy outputs over time, this episode is for you.