MLOps.community

Hamza Tahir, co-founder of ZenML, joins the show to cut through the hype around long-running agents — arguing that at the end of the day, an agent is just a while loop that talks to a model, calls a tool, and writes to a file system. He covers the architecture of agent harnesses (inner and outer), what durable execution actually guarantees (and what it doesn't), and why the ML pipeline paradigm is a cleaner mental model than transactions for most agent workloads.

Hamza also announces Kitaru — ZenML's new open-source execution runtime for async Python agents — built on five years of running ML workloads in enterprise environments.

What we get into:
Agents are while loops: The surprising simplicity under all the tooling: a brain (LLM), hands (tool calls), and a file system, stacked recursively
Inner harness vs outer harness: Why Pydantic AI owns the inner loop while production deployment needs a separate runtime layer
What "long-running" actually means: Why the infrastructure we need to build is about extrapolating the future, not defining a time window today
Durable execution demystified: What checkpointing actually guarantees (infra failures, pod death, network drops) vs. what it never will (external state, bad LLM outputs, Snowflake rollbacks)
ML pipelines vs transactions: Why bursty containers in Kubernetes map more naturally to agent workloads than microsecond-latency queue workers — and why Hamza argues against the complexity tax
Anthropic opening the harness: Why letting other models run Claude Cowork is a "boss move," and what it means for the one-harness vs one-model debate
Human-in-the-loop, done right: The pod-kill-and-resume pattern, and why warm pools matter less when your agent runs for days
Kitaru: ZenML's new open source durable execution runtime: zero-config local, Kubernetes/SageMaker/Vertex in production, built on Pydantic AI integration
Arguing with Claude about Temporal: Hamza's story of spending hours getting an LLM to admit ZenML and Temporal solves the same problem

If you're architecting agents for production, picking between Pydantic AI, LangGraph, and Temporal, or just want to understand what "durable execution" actually means — this is the episode.

// LINKS & RESOURCES
Kitaru on GitHub: https://github.com/zenml-io/kitaru
Kitaru launch blog post: https://www.zenml.io/blog/kitaru-launch
Kitaru on Hacker News: https://news.ycombinator.com/item?id=47520115
Hamza Tahir on LinkedIn: https://www.linkedin.com/in/hamzatahirofficial/
ZenML: https://www.zenml.io/

Timestamps
[00:00] While Loop Checkpointing
[00:24] Long-Running Agents Explained
[01:28] Agent Harness Model Definitions
[06:30] Durability and State Recovery
[11:03] Agent Systems Layers
[18:45] Durability in Agent Systems
[22:07] ML Pipeline vs Transactions
[29:23] Durability vs Guarantees
[33:13] Durability vs Chaos Engineering
[39:50] Kitaru Naming and Purpose
[40:38] Wrap up

#AIAgents #DurableExecution #OpenSource

Rafael (Head of Innovation, iFood) and Daniel (Data and AI Manager, iFood) pull back the curtain on ILO-Agent — iFood's conversational AI ordering system built for 200 million users across Latin America. Recorded live at AI House Amsterdam, this conversation goes deep into the engineering and product decisions behind building recommendation systems and agentic AI, and why the speed of your AI's response might actually be destroying user trust.

The Latency Goldilocks Zone Explained // MLOps Podcast #376 with iFood's Rafael Borger (Head of Innovation) and Daniel Wolbert (Data and AI Manager)

🍕 Recommendation Systems at Scale — Why personalizing for 200M users with wildly different food tastes, budgets, and cultures is a fundamentally different problem than standard ML
🤖 ILO-Agent Deep Dive — What iFood's conversational AI agent actually does, how it handles open-ended requests ("a romantic dinner for two, my wife hates onions"), and where it's headed
⏱️ The Latency Goldilocks Zone — The fascinating insight that LLM responses can be too fast (users don't trust them) or too slow (users abandon) — and how to find the sweet spot
🧠 Perceived vs. Actual Latency — Why showing progress indicators and partial results can make a 6-second response feel instant, and how iFood uses this in production
🛒 The Tinder for Food Experience — How iFood is experimenting with swipe-based discovery to solve "I don't know what I want to eat" for millions of undecided users
🗣️ Voice vs. Text AI Interfaces — Why voice ordering limits you to 6 items in 30 seconds, and why text-based agents need radically different output design
🔗 Agent-to-Agent (A2A) Architectures — What happens when your customer support agent and your ordering agent need to collaborate, and the standardization challenges ahead
📊 Measuring Product-Market Fit for AI — Why the Sean Ellis / Chanel score method breaks down in Brazil, and what iFood uses instead
🏗️ Scalability vs. Ecosystem Health — The real tension between consuming partner APIs aggressively and keeping the food delivery ecosystem sustainable
🌎 Building AI for Global-Local Markets — Why one-size-fits-all AI products fail and how iFood builds for cultural and economic diversity simultaneously.

This episode is for ML engineers, AI product managers, and data scientists building production AI systems at scale — especially if you're working on recommendation, retrieval, or agentic systems in consumer apps.

🔗 Links & Resources
MLOps.community: https://mlops.community
AI House Amsterdam: https://aihouse.amsterdam
iFood: https://www.ifood.com.br/
iFood AILO launch coverage: https://tiinside.com.br/en/10/10/2025/ifood-lanca-ailo-assistente-de-ia-que-inaugura-pedidos-por-conversa/
iFood AI case study (AWS): https://aws.amazon.com/solutions/case-studies/ifood-bedrock/
Related MLOps Community talk — "From Zero to AILO" by Nishikant Dhanuka & Chiara Caratelli: https://home.mlops.community/public/videos/from-zero-to-ailo-lessons-learned-from-building-ifoods-ai-agent-nishikant-dhanuka-and-chiara-caratelli-2025-11-25
ZenML LLMOps database write-up on iFood's hyper-personalized agent: https://www.zenml.io/llmops-database/building-a-hyper-personalized-food-ordering-agent-for-e-commerce-at-scale

⏱️ Timestamps
[00:00] Recommending the unknown
[00:18] Ailo Hyperpersonalization Insight
[06:24] Predictive Personalization Insights
[09:13] "Jet skis" of innovation
[17:45] Consumer Behavior and Chatbots
[26:33] Perceived Latency and Engagement
[33:22] AI-driven UI Evolution
[38:17] LCM Voice Mode Inquiry
[45:20] Chat as Interface
[47:46] Wrap up

Nicolas Alejandro Bogliolo is the AI PM at Despegar, the largest online travel agency in Latin America, and the engineer-product-hybrid behind Sofia, the GenAI travel concierge that beat most of the OTA world to a working multi-agent system.

Before MCP was a standard and before LangChain was widely adopted, his team had already shipped their own orchestration layer and tool protocol in production. This conversation is a rare look at what it takes to build an agentic system that actually books trips, runs on WhatsApp, and keeps adding capabilities without falling over.
Building MCP Before MCP Existed: Inside Despegar's Sofia Agent // MLOps Podcast #375 with Nicolas Alejandro Bogliolo, AI PM at Despegar
What we cover:
- Chappi, the brain of Sofia: how Despegar built an internal orchestration layer when there was nothing off the shelf- Building "MCP before MCP": the custom tool-calling protocol that predated the Anthropic standard- Multi-agent architecture by vertical: flights, hotels, activities, and cars each own their own flow
- Decentralized agent ownership: how any squad in the company can build a flow with central supervision
- Sofia on WhatsApp: making messaging the consumer control center, the way Slack became it for the enterprise
- The five-phase travel arc Sofia covers: dreaming, planning, anticipation, in-trip, and post-trip
- KPI evolution: why "in-scope conversation rate" topped out near 96 percent and what they measure now
- The flight-delay-claim use case and why filing claims through a chatbot is a perfect agent task
- Group trip planning in WhatsApp groups: the next frontier for travel agents
- Sofia as channel of choice: the WeChat-style vision for an agent that handles your entire trip
- Why Despegar held off on giving Sofia the ability to bargain with customers, for now.

Whether you are building production agents, running an OTA, or just curious about how an AI travel concierge actually works under the hood, this episode is full of grounded, in-production lessons from a team that had to invent the patterns the rest of us are now adopting.

Links and Resources:
Despegar: https://www.despegar.com
Sofia announcement: https://investor.despegar.com/news-presentations/news-releases/news-details/2024/Despegar-revolutionizes-the-tourism-industry-introducing-the-regions-first-Generative-AI-Travel-Assistant
Sofia coverage on PhocusWire: https://www.phocuswire.com/despegar-debuts-genai-travel-assistant-remembers-previous-interactions
MLOps Community: https://mlops.community
Subscribe for more agent and AI infra deep dives

Timestamps
[00:00] Sophia Travel Concierge AI
[00:38] Sophia Multi-Agent System
[06:00] AI Limitations in Practice
[13:52] Travel Planning Exploration
[18:03] Group Travel Decision Making
[21:32] Agent Ecosystem Design
[30:14] Sofia's Travel Assistant Vision
[33:35] Orchestration and MCP Design
[40:13] Sophia Negotiation Concerns
[40:47] Wrap up

#AIAgents #MCP #AgenticAI

This episode is brought to you by the MLflow team. Check out more information at MLflow.org.

What does it actually take to build voice AI at a billion-interaction scale? This episode features an ex-Amazon voice AI engineer who built customer support systems handling 2 billion+ interactions — now working on next-gen voice agent platforms. Anurag digs deep into the real engineering tradeoffs, design patterns, and use cases that separate production-grade voice agents from demos.

Voice Agent Use Cases // MLOps Podcast #374 with Anurag Beniwal, Member of the Technical Staff at ElevenLabs

🎙️ Topics covered:
🔹 Cascaded vs. speech-to-speech — Why cascaded systems still win in production, and how to make them feel natural without sacrificing control
🔹 Latency masking — Foreground/background model architecture and how to buy yourself time while deep retrieval runs
🔹 Constellation of models — Using Haiku for tool calling, fine-tuned smaller models for response generation, and why "one model for everything" breaks at scale
🔹 Turn-taking & ASR challenges — Why voice is harder than chat: accents, noise, silence detection, and domain-specific fine-tuning
🔹 Level 1 vs Level 2 customer support — Why today's agents max out at Level 1 and what it takes to capture Level 2 expert judgment
🔹 Inbound vs. outbound sales agents — Where voice agents are already winning, and why inbound lead qualification beats cold outbound
🔹 Booking, reservations & concierge — The clearest near-term wins for voice agents across hospitality, home services, and SMBs
🔹 Continual learning from natural language feedback — How to build agents that improve from real operator feedback without ML expertise
🔹 Conversational TTS — Why passing full conversation history to your TTS model changes everything for tone consistency
🔹 User tiers for voice platforms — Non-technical business owners vs. developers vs. enterprise: why one interface doesn't fit all.

If you're building production voice agents, evaluating voice AI vendors, or scaling AI-first customer support — this episode is packed with hard-won lessons from someone who's done it at Amazon scale.

🔗 Links & Resources:
MLOps.community: https://mlops.communityGoogle Scholar: https://scholar.google.com/citations?user=g_QB5WgAAAAJ&hl=en&o
Amazon science page: https://www.amazon.science/author/anurag-beniwal
Join the Community: https://go.mlops.community/YTJoinIn
Get the newsletter: https://go.mlops.community/YTNewsletter
MLOps GPU Guide: https://go.mlops.community/gpuguide

⏱️ Timestamps
[00:00] Cascaded Systems Control Challenge
[05:35] Voice vs Chat Complexity
[14:16] MLflow's open source platform
[15:03] AI Model Constellations
[23:00] Model Constellations Use Cases
[31:40] Voice vs Text Context
[33:54] Voice as Thought Capture
[42:11] Cascaded vs Speech-to-Speech Debate
[50:02] Wrap up

Jesse Vincent is the Founder & CEO of Prime Radiant and creator of Superpowers — the most-used Claude Code plugin in the world. He built the first agentic software development methodology from scratch while managing MIT interns in the early 2000s, and hasn't written a line of code manually since October.

The Creator of Superpowers: Why Real Agentic Engineering Beats Vibe Coding // MLOps Podcast #373 with Jesse Vincent, Founder & CEO of Prime Radiant

In this conversation, Jesse walks Demetrios through the full Superpowers system: why he thinks most developers are still approaching agentic coding wrong, how he designs skills that force LLMs to stop rationalizing and actually follow rules, and what he's building next at Prime Radiant — including Green Field, an unreleased tool for reverse-engineering legacy codebases into specs. This one is for developers who want to go beyond "vibe coding" and build AI-assisted workflows that actually scale.

🔧 Topics Covered
🧠 The Superpowers Methodology — How the brainstorming skill extracts what you actually want before you hand work to an agent, and why most developers skip this step
📋 Spec-Driven Development & Plan Files — Why Jesse insists on TDD, DRY, and YAGNI for every agentic task, and how planning skills generate per-task context blocks agents can actually execute on
🐛 Debugging with Agents — Jesse's systematic approach to root cause analysis, reproduction cases, and the 30 years of debugging instinct he's baked into a skill
🔄 Pressure Testing LLM Skills — How Claude fires up sub-agents and stress-tests its own rules to catch rationalization before it shows up in production
🛠️ Clearance IDE — Jesse's new Markdown-native development environment built for humans working alongside AI, with a history pane for file navigation
📦 Green Field (Unreleased) — A toolset for turning old codebases or built products into clean specs — not yet public but dropping soon from Prime Radiant
🧑‍💼 Management as the Magic Trick — Why the real unlock of tools like Superpowers is that they make every developer a manager, and why that transition is hard the first time
⚖️ Software Ethics in the Agent Era — Reverse engineering, license washing, open source cloning, and whether the value of software itself is collapsing

🔗 Links & Resources
Prime Radiant: [https://prime-radiant.com](https://prime-radiant.com/)
Superpowers on GitHub: https://github.com/prime-radiant-inc
Clearance IDE: https://github.com/prime-radiant-inc (check repo)
MLOps.community Slack: https://go.mlops.community/slack
MLOps.community website: [https://mlops.community](https://mlops.community/)

⏱️ Timestamps
[00:00] Greenfield Toolset Insights
[00:27] Superpowers Kit Evangelism
[08:06] Hyperbolic's GPU Cloud
[17:48] Debugging Skill Creation
[22:12] Skill Extraction Strategy
[31:15] Smallest Harness
[41:06] Software supply chains
[48:56] Visual Precision Challenges
[54:09] Creative Feedback Loops
[1:04:24] MLflow's Gen AI
[1:05:55] Wrap-up