Union.ai buying intent

The agent builders I met at the MLOps Seattle AI Agents Summit weren’t asking about models. They were asking about boring infra. Across booth conversations, the pattern was impossible to miss. I talked to a sales engineer from a compliance company building an agent and looking for a platform to run it on. A founder working on voice models for internal training. A team doing edge inference for fleet management. Someone in healthcare experimenting with agent deployment. A retail engineering lead trying to plug agents into existing workflows. Different industries. Different use cases. Different definitions of “agent.” But within about 90 seconds, almost every conversation landed on the same questions: * How do you handle durability if an agent crashes mid-loop? * What’s your caching story? * How fast do tasks actually start? * Can this run at the edge, on-prem, or in regulated environments? * How do retries work when a failure happens three levels deep in a tool call? Not one person asked which model we recommend. Nobody asked about prompting. Nobody asked about benchmarks. The “agents” conversation has quietly become a distributed systems conversation, and a lot of the industry still hasn’t caught up. A few things feel clear: 1. The bottleneck has moved. A year ago the interesting question was: can the model do the task? Now it’s: can the system around the model do it reliably, cheaply, and fast enough for a human or another agent to use? 2. Cold boot is the new latency. If a task takes 45 seconds to start, the agent feels broken. Cold start performance matters way more than people think. 3. Caching is not an optimization. When every step may call an LLM, a search API, a retrieval layer, and a tool, recomputing deterministic work gets expensive fast. 4. Durability beats cleverness. The agents that survive production aren’t the ones with the fanciest planner. They’re the ones that can fail, recover, and resume exactly where they left off. The interesting open source problems are infra problems. The OSS conversations I had weren’t about which model to fine-tune. They were about checkpointing, resuming pipelines, edge-friendly containers, and not wasting GPU-hours. None of this is flashy. It’s the layer that makes a 20-second demo look the same as a 20-minute production workflow. But if you’re trying to understand where the real product and platform opportunities are for 2026, listen to what builders are complaining about, not what keynotes are hyping. The boring layer underneath the agent is the product. Are you seeing the same shift in your own work?

This week we'll be building with Gemma 4! Gemma 4 is Google DeepMind's latest open model family, built from Gemini 3 research and released under Apache 2.0. It ships in four sizes from Effective 2B up to 31B Dense, with native function calling, configurable thinking modes, and multimodal input across a 256K context window. Plenty of surface area for agentic workflows and fine-tuning experiments. 📅 Join the event for the live stream: https://luma.com/0j0wzgaz 💬 Join the conversation in Slack: https://slack.flyte.org/ I've testing it's visual understanding for scene comprehension and object detection. Very impressed so far, and I'm excited to test on the smaller versions to see how they do!