The last two weeks brought major shifts in how teams deploy, scope, and govern production agents. These posts trace a clear pattern: from infrastructure abstraction to architectural rethinking to governance that follows users, not prompts.
Managed Deep Agents is now in Public Beta
LangSmith's managed agent platform removes the infrastructure burden from production deployments. Instead of stitching together durable execution, sandboxes, memory, and channel integrations, teams can ship agents with a single command while retaining full control over models, prompts, and tools. This is the "Vercel for agents" moment: abstraction that doesn't lock you into opinionated behavior, just handles the runtime complexity you'd otherwise rebuild yourself.
// Local development to production with one deploy
langsmith deploy agent \
--name "github-pr-reviewer" \
--channels slack,github \
--sandbox enabledBuilding monday.com Sidekick: why capable agents need more than just tools
Monday.com's team learned that adding more tools to a single agent degrades performance in production. Their solution: bounded subagents with domain-specific reasoning, tools for controlled operations, and sandboxes for iterative work like file processing. This architectural pattern cuts through the "universal agent" trap and replaces it with composable specialization. Context bloat and combinatorial test explosion both disappear when you stop treating every capability as another tool in one giant registry.
Meta Open-Sources Muse Glimmer: A 30B Local Agentic Model Optimised for On-Device Execution
Meta released Muse Glimmer, a 30B parameter agent model that runs on consumer GPUs (24-32 GB VRAM) through 4-bit quantization and speculative decoding. At 17-20 GB memory footprint with 3.1x throughput improvements, it's optimized for autonomous tool use and coding workflows without cloud dependencies. Apache 2.0 license and native support in llama.cpp, Ollama, and MLX make this immediately usable for teams that need privacy-preserving agents or want to avoid variable cloud token costs.
Why managed agents are the next big thing in agent building
The evolution from frameworks (write your own everything) to managed platforms (bundled harnesses plus infrastructure) mirrors what happened with serverless. Standardized control mechanisms like AGENTS.md and MCP now combine with built-in runtime execution, sandboxes, eval, and auth. Teams stop assembling production infrastructure from scratch and start focusing on behavior and business logic. The shift is from "build the platform" to "configure the runtime."
How to ground Genie Agents in both structured data and documents without losing governance
Databricks places governance at the data layer, not the model layer. Genie Agents run with end-user credentials, so Unity Catalog's row filters, column masks, and ABAC policies automatically scope every query. Different users asking identical questions get different, correctly-filtered answers based on their permissions. No per-user prompt engineering, no brittle policy logic in the agent layer—just SQL-level access control that already works for your tables and now extends to documents in Unity Catalog Volumes.
The throughline: managed infrastructure is abstracting deployment complexity, architectures are shifting from monolithic to bounded, and governance is moving to the data perimeter where it belongs. Production agents are becoming cheaper to run, easier to scope, and safer to ship. 🚢