Agents move from idea to stack

A multi-agent system is a setup where one model acts like a manager and passes parts of a job to specialist models, and that design is moving from demos into product stacks. (docs.langchain.com) The basic patterns are now documented in vendor and framework guides: routers classify a task and send it to the right worker, handoffs switch control based on state, and subagents run subtasks in parallel before an orchestrator combines the results. (docs.langchain.com) (microsoft.github.io) (developers.openai.com) LangChain’s multi-agent guide says developers use these systems for context management, parallelization, and distributed development, and warns that some jobs still work with one agent plus tools. Microsoft’s reference architecture adds patterns such as a semantic router with a lighter classifier first and a dynamic agent registry for runtime discovery. (docs.langchain.com) (microsoft.github.io) That shift showed up in infrastructure on April 11, when NVIDIA published a post positioning MiniMax M2.7 for “agentic harnesses” and long-running assistants on its platform. NVIDIA said the open-weights model is available through its stack and the broader open-source inference ecosystem. (developer.nvidia.com) MiniMax M2.7 is a sparse mixture-of-experts model, which means it contains 230 billion total parameters but activates 10 billion per token by routing work to a small subset of 256 experts. NVIDIA’s post lists a 200,000-token context window and says the design is meant to keep inference costs lower than a dense model of similar size. (developer.nvidia.com) NVIDIA said it added performance work for the M2 series in vLLM and SGLang, including a query-key root mean square normalization kernel and a TensorRT-LLM floating point 8 mixture-of-experts kernel. NVIDIA said those changes target the communication and memory bottlenecks that show up when large expert models run on graphics processing units. (developer.nvidia.com) The company also tied the model to NVIDIA NemoClaw, an open-source reference stack for always-on assistants, and to OpenShell, a runtime it describes as a secure environment for autonomous agents. NVIDIA said developers can launch that setup on the NVIDIA Brev cloud graphics processing unit platform. (developer.nvidia.com) MiniMax’s own repository pitches M2.7 as a model for “complex agent harnesses,” “Agent Teams,” and dynamic tool search, and lists benchmark claims including 56.22 percent on SWE-Pro and 57.0 percent on Terminal Bench 2. The repository also says an internal version of M2.7 improved a programming scaffold over more than 100 rounds and delivered a 30 percent performance gain. (github.com) OpenAI’s Codex docs show the same ideas landing in developer tools: Codex can spawn subagents in parallel, wait for all of them, and return one combined answer, while exposing thread controls and approval handling in the command line interface. The docs also say subagent runs cost more tokens than comparable single-agent runs because each child agent does its own model and tool work. (developers.openai.com) The result is a stack taking shape in layers: documented orchestration patterns at the top, agent runtimes in the middle, and model and kernel tuning underneath. The new pitch is less about whether agents exist and more about which routing, handoff, and parallel-worker pattern a team can ship. (docs.langchain.com) (microsoft.github.io) (developer.nvidia.com)