--- title: "Workflows That Outlive Your Laptop" newsletter: "MLOps Community" date: 2026-03-19 source: https://aaif.live/newsletters/mlopscommunity/2026-03-19-workflows-that-outlive-your-laptop --- # Workflows That Outlive Your Laptop *MLOps Community — Agentic AI Foundation, 2026-03-19* More great write-ups from our Coding Agents conference, one on the mindset shift on working with agents [https://www.linkedin.com/pulse/do-outsource-thinking-michael-stricklen-khwjc/] and a deeper look at Dexter Horthy’s talk about disciplined workflows like RPI [https://betterquestions.ai/the-necessary-evolution-of-research-plan-implement-as-an-agentic-practice-in-2026/]. If you want talks so good they inspire you to write, join us in Seattle on April 14 [https://luma.com/ai-agents-summit-seattle]. ## Working as Intended The next wave of agent outages will not come from hallucinations. They will come from distributed systems behaving exactly as designed. What keeps you awake more? Model errors or System failures [MODEL](https://gatewaze.mlops.community/offer/surveys/?sid=yesno&question=The+next+wave+of+agent+outages+will+not+come+from+hallucinations.+They+will+come+from+distributed+systems+behaving+exactly+as+designed.+What+keeps+you+awake+more%3F+Model+errors+or+System+failures&y=MODEL&n=SYSTEM&oneclick=true&accept=true) [SYSTEM](https://gatewaze.mlops.community/offer/surveys/?sid=yesno&question=The+next+wave+of+agent+outages+will+not+come+from+hallucinations.+They+will+come+from+distributed+systems+behaving+exactly+as+designed.+What+keeps+you+awake+more%3F+Model+errors+or+System+failures&y=MODEL&n=SYSTEM&oneclick=true&accept=true) ## Open for business Control, cost, and portability beat the convenience of a closed API when shipping production systems. ## Ship Agents: Free Virtual Conference - March 26 Over two hours of production agent talks from practitioners at Amazon, Nvidia, Microsoft, AWS, IBM, and early-stage teams building in the same trenches. The lineup is heavy on failure modes, fixes, and the operational side of agents that most conference talks skip over. Including: * Divya Mahajan (Amazon) - eight silent failure patterns in multi-turn agents and how teams detect them * Pratik Mehta (Nvidia) - 12 months running a production sales agent at scale, and what had to be torn out and rebuilt * Xia Hua (Traceforce) - pentesting MCP servers and real vulnerabilities found in widely used implementations * Elizabeth Fuentes (AWS) - context engineering patterns to stop agents choking on large tool outputs * Deepak Kamboj (Microsoft) - why teams are moving from raw autonomy to deterministic workflows and guardrails All free and online - March 26, 9:00 AM PT / 5:00 PM CET REGISTER FOR FREE [REGISTER FOR FREE](https://home.mlops.community/home/events/ship-agents-a-virtual-conference-zzr92e4gsi?agenda_day=69b2b893d6b7fbe3b94ffdbf&agenda_track=69b2b893d6b7fbe3b94ffdd6&agenda_stage=69b2b893d6b7fbe3b94ffdc5&agenda_filter_view=stage&agenda_view=list) [https://home.mlops.community/home/events/ship-agents-a-virtual-conference-zzr92e4gsi?agenda_day=69b2b893d6b7fbe3b94ffdbf&agenda_track=69b2b893d6b7fbe3b94ffdd6&agenda_stage=69b2b893d6b7fbe3b94ffdc5&agenda_filter_view=stage&agenda_view=list](https://home.mlops.community/home/events/ship-agents-a-virtual-conference-zzr92e4gsi?agenda_day=69b2b893d6b7fbe3b94ffdbf&agenda_track=69b2b893d6b7fbe3b94ffdd6&agenda_stage=69b2b893d6b7fbe3b94ffdc5&agenda_filter_view=stage&agenda_view=list) ## Curated finds to help you stay ahead ## Parallel Agents, Parallel Problems Twenty agents rewriting a monorepo overnight will surface every infra shortcut you've been living with. Isolation gaps, storage bottlenecks, no observability, no rollback - they all show up at once. The stack has four layers that matter. Each one has a clear failure mode if you skip it. Execution environments The first thing that breaks is isolation. An agent running in a shared environment will step on other agents' processes, leave behind conflicting state, or - if it's executing generated or user-supplied code - create security exposure you didn't plan for. One team building a parallel code review system found agents intermittently failing with no clear error. The cause: filesystem side effects from one agent's build process interfering with another's test run. No isolation, no clean signal. Each agent needs its own ephemeral compute environment, spun up fast and torn down cleanly: * Containers are the default choice. Cold start times add up at scale and resource limits need to be explicit, but the tooling is mature. * Micro-VM approaches (Firecracker, gVisor) get you stronger isolation boundaries at close to container startup speeds - worth it if agents are running untrusted code. * E2B [https://e2b.dev] is a good default if you don't want to manage this layer yourself - purpose-built secure cloud sandboxes for AI agent workloads, with fast provisioning and a simple API. Once execution is isolated, the next problem is what agents actually read and write. Filesystem and storage This is where the standard cloud options break down in ways that aren't obvious until you're already in trouble: * EBS is single-attach - useless for parallel agents that need shared access * S3 lacks POSIX semantics and has the wrong latency for constant small reads and writes * EFS handles concurrent access but write performance degrades hard under the kind of workload coding agents produce - rapid generation, frequent checkpointing, parallel builds The underlying tension: agents want local-disk semantics. Your bill and your reliability want object-store behaviour. The options below are different attempts to close that gap: * Archil [https://archil.com] - managed shared volumes backed by S3. Up to 30x lower latency than direct S3 access, up to 100x faster writes than EFS. Agents mount shared underlying storage while working in isolated volume contexts. Managed, so no metadata infrastructure to run. * JuiceFS [https://juicefs.com] - open source. Object storage as the data layer, separate metadata engine (Redis, TiKV, MySQL), POSIX interface on top. Widely used in ML training infra. You control the performance profile by choosing your metadata backend, but that backend is yours to operate. Pick Archil if you want managed. Pick JuiceFS if you want to own the knobs. Once storage is stable, the question is how agents interact with versioned code without creating branch chaos. Code versioning and agent deployment Background agents can't just clone a repo and run Git commands. At scale that pattern produces collisions, corrupted history, and branches that nobody can track. You need agents to commit and branch programmatically, with output that's reviewable before it touches main. * Mesa [https://mesa.dev] - programmatic Git hosting built for machine-driven workflows. Agents create commits, manage branches, and read files via API, without cloning or maintaining local Git state. Designed for high-throughput workloads across many repos. * Terminal Use [https://terminaluse.com] - deployment infrastructure for background agents, with Git-native branching and rollback. Agent work happens on a branch, output is reviewable, rollback is built in rather than manual. YC-backed, early stage. Observability With a human engineer, you get implicit logging through commits, PR comments, and ticket updates. With background agents, that signal disappears unless you build it back in deliberately. What specifically breaks: an agent that stalls mid-task produces no output and no error, so you have no idea whether it's working, stuck, or silently corrupting state. An agent that succeeds on the wrong interpretation of a task looks fine in logs until someone reviews the output hours later. Token burn and storage I/O spike with no obvious cause. Standard metrics - CPU, memory, request latency - don't surface any of this. Langfuse, Arize, and others are adding agent-specific tracing support, but it's still catching up with how agents behave in practice. Treat observability as a design requirement, not a retrofit. What to figure out before you scale The three cost levers are: how long agents run, how often they branch, and how much high-performance shared storage they use. A fleet of agents running multihour background tasks on fast shared volumes generates a bill that's easy to miss if you sized the budget against single-agent usage. Model it early. Before you move agents into production: decide your execution substrate, choose a shared filesystem approach, and have a plan for how agents version their output. Those three decisions shape everything downstream - reliability, cost, and whether you can trust what agents produce at scale. ## Durable Execution and Modern Distributed Systems Your week-long agent job doesn’t have to die because one flaky API timed out on Tuesday. This discussion breaks down durable execution as a way to keep long-running workflows, agent loops, and async processes moving without turning every failure into a manual recovery job. * It separates business logic from reliability logic, so retries, recovery, and state persistence stop cluttering application code. * The key programming split is simple: workflows control execution, activities handle I/O, which makes agent loops easier to reason about. * It also goes beyond coarse checkpointing by replaying from durable state, even across crashes, sleeps, and long human pauses. That matters when your agent has to keep running long after your laptop, container, or cloud region gives up. [https://podcasts.apple.com/gb/podcast/durable-execution-and-modern-distributed-systems/id1505372978?i=1000755798178](https://podcasts.apple.com/gb/podcast/durable-execution-and-modern-distributed-systems/id1505372978?i=1000755798178) [https://home.mlops.community/home/videos/durable-execution-and-modern-distributed-systems](https://home.mlops.community/home/videos/durable-execution-and-modern-distributed-systems) [https://open.spotify.com/episode/4JM4Z4F91918xvLzilS7zd?si=24a4d87201d54382](https://open.spotify.com/episode/4JM4Z4F91918xvLzilS7zd?si=24a4d87201d54382) ## mAIdAI: Building a Personal Assistant with Google Cloud and Vertex AI You do not need a giant agent stack to remove daily workflow drag. This post shows a personal assistant built for one person’s docs, shortcuts, and habits, using Google Chat as the front end and a small serverless backend to route requests by cost, latency, and complexity. * Quick commands skip the model entirely for links, snippets, and other fixed responses, which keeps common tasks fast and cheap. * Slash commands add lightweight prompt templates for repeatable jobs like debugging or rewriting without building a full UI. * The backend stays minimal with FastAPI, Cloud Run, and Vertex AI, using a personal context file to ground every response. The useful bit is not the chatbot wrapper but the routing model that decides when an LLM is needed and when it is not. [Read the blog](https://home.mlops.community/home/blogs/maidai-building-a-personal-assistant-with-google-cloud-and-vertex-ai) ## Chaigent: An affordable alternative to Gemini Enterprise on Google Cloud Flat per-seat pricing can make agent platforms hard to justify when usage is sporadic. This write-up shows a DIY stack that swaps a managed enterprise frontend for open-source tooling while still using hosted reasoning services underneath. * A serverless UI built with Chainlit handles sessions, auth, and streaming responses without licensing overhead. * Vertex AI Agent Engine remains the reasoning layer, with Cloud SQL storing chat history for ownership and analytics. * The trade-off is clear: lower cost and full customization in exchange for building governance, security, and ops yourself. It highlights how architecture choices, not model choice alone, often decide whether agents scale economically. [Read the blog](https://home.mlops.community/home/blogs/chaigent-an-affordable-alternative-to-gemini-enterprise-on-google-cloud) ## IN PERSON EVENTS * Munich [https://luma.com/nfwzg8wv] - March 24 * NYC [https://luma.com/ainativestack] - March 24 * London [https://luma.com/h57hqyvq] - March 26 * Lagos [https://luma.com/gcgmri06] - March 28 * Seattle [https://luma.com/ai-agents-summit-seattle] - April 14 ## VIRTUAL EVENTS * Coding Agents Lunch & Learn [https://home.mlops.community/home/events/coding-agents-lunch-and-learn-session-6-d8o6tqqru7] - March 20 * Operationalizing AI Agents: From Experimentation to Production [https://home.mlops.community/home/events/operationalizing-ai-agents-from-experimentation-to-production-ps4tmp6g3o?agenda_day=698b7637ad917410bb9de954&agenda_track=698b7638ad917410bb9de96c&agenda_stage=698b7638ad917410bb9de95a&agenda_filter_view=stage&agenda_view=list] - March 25 * Ship Agents: A Virtual Conference [https://home.mlops.community/home/events/ship-agents-a-virtual-conference-zzr92e4gsi?agenda_day=69b2b893d6b7fbe3b94ffdbf&agenda_track=69b2b893d6b7fbe3b94ffdd6&agenda_stage=69b2b893d6b7fbe3b94ffdc5&agenda_filter_view=stage&agenda_view=list] - March 26 ## Seed of Doubt [https://forms.gle/8EDvXGizxyFVKfwy8](https://forms.gle/8EDvXGizxyFVKfwy8) --- Source: https://aaif.live/newsletters/mlopscommunity/2026-03-19-workflows-that-outlive-your-laptop