Teo Richie
Richie is a Lead Experience Engineer based in Kuala Lumpur, Malaysia, with five years of experience building software products. He works at the intersection of backend systems and user-facing product, with a focus on making complex infrastructure feel simple to use. His recent work includes Alfred, a multi-agent orchestration platform. This project pushed him deep into the practical challenges of running AI systems in production, which is what this talk is about.
Intervention
Everyone talks about prompts, models, and RAG. Nobody talks about the 80% of work that actually gets AI agents running in production.
This talk is about the integration layer. The OAuth flows, channel routing, credential management, scheduling infrastructure, and observability tooling that separates a working demo from a system your users can rely on every day.
I'll draw from a system I've participated in building: Alfred, a multi-agent orchestration platform. We'll start with integrations. An agent is only as useful as what it can reach. Connecting to Gmail, Slack, Google Calendar, Microsoft 365, or Telegram sounds straightforward until you're dealing with token expiry, scope changes, and multiple users with different credentials. We'll look at how to build a credential layer that handles all of this cleanly, so agents can act on behalf of users without the whole thing falling apart when a token rotates.
From there we'll talk about routing. In a real deployment, messages come in from multiple channels simultaneously. Telegram, Slack, and web chat all have different UX contracts and different audiences. The answer isn't to write channel-aware agent code. It's to build a routing layer that maps channels to agents declaratively, so you can change behavior without touching your agent logic.
We'll also cover scheduling. Agents are far more powerful when they run on a schedule rather than waiting to be asked. We'll look at how to treat cron jobs as first-class agent invocations, with full run history and cost tracking so you know exactly what every scheduled run is doing and what it's spending.
Finally we'll talk about observability. If you can't see what your agent is doing in real time, and you can't track what it's costing you, you can't trust it in production. We'll look at patterns for tracking subagent spawns, monitoring delegation health, and surfacing live activity so problems surface before your users notice them.
By the end of this talk you'll have a clear mental model of the integration layer and a set of concrete, framework-agnostic patterns you can take back to whatever stack you're building on.