--- title: "MLflow Meets Long-Running Agents" newsletter: "MLOps Community" date: 2026-01-22 source: https://aaif.live/newsletters/mlopscommunity/2026-01-22-mlflow-meets-long-running-agents --- # MLflow Meets Long-Running Agents *Plus disposable code, GPU tuning, low-latency features, and semantic IDs* *MLOps Community — Agentic AI Foundation, 2026-01-22* The digital landscape of OpenAI continues to pivot, with a game-changing move that's all about unlocking its insightful takeaways they’ve announced that ads are coming to ChatGPT [https://openai.com/index/our-approach-to-advertising-and-expanding-access/]—and here’s the kicker: it’s not the only thing coming, as Blockbuster now has extra copies of Twister in, and are offering 3 vids for 2 with a free 2L bottle of Pepsi Crystal. Head to your local store now. ## Total Recall or Total Regret? Long-term agent memory is mostly a liability in disguise. Do you default to forgetting or remembering? [FORGETTING](https://gatewaze.mlops.community/offer/surveys/?sid=yesno&question=Long-term+agent+memory+is+mostly+a+liability+in+disguise.+Do+you+default+to+forgetting+or+remembering%3F&y=FORGETTING&n=REMEMBERING&oneclick=true&accept=true) [REMEMBERING](https://gatewaze.mlops.community/offer/surveys/?sid=yesno&question=Long-term+agent+memory+is+mostly+a+liability+in+disguise.+Do+you+default+to+forgetting+or+remembering%3F&y=FORGETTING&n=REMEMBERING&oneclick=true&accept=true) ## A clear outcome Most said they tracked outcomes rather than throughput. ## What the next generation of AI stacks gets right AI agents don’t fail because LLMs aren’t smart enough. They fail because context is fragmented. When context lives across APIs, databases, and pipelines, you get brittle retrieval, rising costs, and agents that work in demos—but fall apart in production. The fix isn't another framework. It's the context engine. Join Simba Khadder, Cofounder of FeatureForm and Head of Redis’ context engine, on January 28 to learn how a single layer for data, memory, and search makes AI systems reliable, observable, and scalable. Sign up here [Sign up here](https://events.redis.io/wb-predictions-2026?utm_source=mlops-community&utm_medium=cpa&utm_campaign=2026-01-ai_in_production-newsletter&utm_content=wb-2026-01-28-all_you_need_is_context-amer-701N100000ee0f7) [https://events.redis.io/wb-predictions-2026?utm_source=mlops-community&utm_medium=cpa&utm_campaign=2026-01-ai_in_production-newsletter&utm_content=wb-2026-01-28-all_you_need_is_context-amer-701N100000ee0f7](https://events.redis.io/wb-predictions-2026?utm_source=mlops-community&utm_medium=cpa&utm_campaign=2026-01-ai_in_production-newsletter&utm_content=wb-2026-01-28-all_you_need_is_context-amer-701N100000ee0f7) ## Curated finds to help you stay ahead ## MLflow Leading Open Source MLflow’s open source work is starting to look very different in the agent era. The core ideas are familiar, but the failure modes are not. * Most production “agents” are still chatbots, but they now call tools, write data, and fail across long, multi-turn conversations. * Evaluation is shifting from single responses to session-level judges that catch repetition, inconsistency, and user frustration. * Feedback, memory, and governance are becoming first-order concerns as agents touch sensitive data and real systems. It’s a clear view of how open source ML tooling is being reshaped to handle agents that behave nothing like static models. [https://podcasts.apple.com/us/podcast/conversation-with-the-mlflow-maintainers/id1505372978?i=1000745468976](https://podcasts.apple.com/us/podcast/conversation-with-the-mlflow-maintainers/id1505372978?i=1000745468976) [https://home.mlops.community/home/videos/mlflow-leading-open-source](https://home.mlops.community/home/videos/mlflow-leading-open-source) [https://open.spotify.com/episode/1UOyLjzvgeMrJzOR10xKIz?si=c0a33a80e32c4f00](https://open.spotify.com/episode/1UOyLjzvgeMrJzOR10xKIz?si=c0a33a80e32c4f00) ## How Universal Resource Management Transforms AI Infrastructure Economics AI teams are scrambling for GPUs while usable compute sits idle. The conversation flips the usual hardware story by arguing that inference workloads rarely need the latest accelerators, and that waiting months for GPUs is often the slowest option. * Inference bottlenecks are memory, not raw compute, making CPU-heavy systems with large RAM surprisingly effective. * Technologies like CXL let teams extend memory cheaply, but tooling and CUDA-first ecosystems slow adoption. * Smaller models and tightly scoped agents shift workloads toward many modest machines instead of a few massive ones. Shipping sooner often means using the hardware you already have, then upgrading only when the workload proves it. [https://podcasts.apple.com/gb/podcast/how-universal-resource-management-transforms-ai-infrastructure/id1505372978?i=1000745930568](https://podcasts.apple.com/gb/podcast/how-universal-resource-management-transforms-ai-infrastructure/id1505372978?i=1000745930568) [https://home.mlops.community/home/videos/how-universal-resource-management-transforms-ai-infrastructure-economics](https://home.mlops.community/home/videos/how-universal-resource-management-transforms-ai-infrastructure-economics) [https://open.spotify.com/episode/3kbiEup5QrPYYf0mKK3SSQ?si=c3e0467ea78e492b](https://open.spotify.com/episode/3kbiEup5QrPYYf0mKK3SSQ?si=c3e0467ea78e492b) ## Upcoming Webinar: Serving LLMs in Production with Cast AI ## VIRTUAL EVENTS On Feb 5, we’re hosting a live session with Cast AI on what breaks once LLMs leave demos. We’ll focus on concrete questions teams are dealing with right now: * why inference and agent workloads cause cloud costs to spike * when managed APIs stop making sense and teams move to Kubernetes * how latency and scaling behave with multi-model and agentic setups There’s time for live Q&A, so you can ask about your own architecture, cost trade-offs, or scaling limits. [Register here](https://home.mlops.community/home/events/serving-llms-in-production-performance-cost-and-scale-wae4fhiyqq?agenda_day=6964f96e50a403eb35751ec2&agenda_track=6964f96e50a403eb35751efb&agenda_stage=6964f96e50a403eb35751ec8&agenda_filter_view=stage&agenda_view=list) ## Upcoming Reading Group: Agent Use for Coding in 2025 Our first Reading Group of 2026 covers Professional Software Developers Don’t Vibe, They Control: AI Agent Use for Coding in 2025 [https://arxiv.org/abs/2512.14012]. It looks at how developers are actually using AI agents in real workflows - less “vibe coding,” more control, orchestration, and reliability. Led by @Anna Yoon, @Valdimar Eggertsson, @Rohan Prasad, and @Lucas Pavanelli, the session connects the paper’s findings to day-to-day developer work and production systems. [Register here](https://home.mlops.community/home/events/mlops-reading-group-january-agent-use-for-coding-in-2025-rfq9erfz1n) ## IN PERSON EVENTS * Barcelona [https://luma.com/tsnwhu9d] - January 29 * San Francisco [https://luma.com/1lyxx9en] - February 9 ## A gift for the new guy [https://forms.gle/u6HcHv1mRVczMw4M8](https://forms.gle/u6HcHv1mRVczMw4M8) --- Source: https://aaif.live/newsletters/mlopscommunity/2026-01-22-mlflow-meets-long-running-agents