AI Agents

AI agents put different pressure on GPU infrastructure than a chat completion does: longer sessions, tool-calling loops that resend context on every step, and often several models running at once, a planner, executors, embedders. This cluster covers the infrastructure that agent frameworks need but rarely document well.

Framework-specific deployment guides cover LangGraph, CrewAI, A2A multi-agent systems, Microsoft Agent Framework, SmolAgents, and open-source coding agents like OpenHands. Infrastructure posts go deeper on the pieces that make agents work in production: persistent memory (Mem0, Zep), code-execution sandboxes (E2B, Daytona, Firecracker), workflow orchestration (Temporal, Inngest, Restate), and MCP server deployment. We also cover evaluation, running SWE-bench, GAIA, and BFCL tool-calling benchmarks at scale, and reinforcement learning specifically for agents, since agent RFT trains on multi-step trajectories rather than single-turn text.

The pillar post, Agentic RAG on GPU Cloud, covers the stack most agent deployments actually need: embedding, vector search, and LLM inference colocated to hit sub-200ms time-to-first-token. Two case studies, 100 concurrent agents on bare metal and a sub-200ms RAG pipeline serving 2M queries a day, show what this looks like at real scale.