complete-ai-agent-stack-deployment-cloud-managed-from-scratch
Sequences a complete, end-to-end AI agent stack deployment built on managed cloud services from scratch — a cloud landing zone, agent control-flow architecture, an LLM gateway routing across managed provider APIs (vendor-neutral: Anthropic/OpenAI/Azure OpenAI/Bedrock/ Vertex AI), a managed vector database for RAG, MCP servers for tool access, an evaluation-and-guardrails harness, and cost/latency monitoring. This is an integration/orchestration skill that sequences several existing tool-specific skills in the correct order and flags the handoff points between them — it does not restate their internals. Use when a user asks to "build a production AI agent stack using managed LLM APIs from scratch," "stand up an agent platform with a managed vector database and MCP tools," "give me the end-to-end sequence for a cloud-managed agent deployment," or "design the full pipeline from cloud account to a production agent with evals and cost monitoring."
selvarajmurugesan90 · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add selvarajmurugesan90/ops-engineering-skills --skill complete-ai-agent-stack-deployment-cloud-managed-from-scratch技能说明
name: complete-ai-agent-stack-deployment-cloud-managed-from-scratch description: > Sequences a complete, end-to-end AI agent stack deployment built on managed cloud services from scratch — a cloud landing zone, agent control-flow architecture, an LLM gateway routing across managed provider APIs (vendor-neutral: Anthropic/OpenAI/Azure OpenAI/Bedrock/ Vertex AI), a managed vector database for RAG, MCP servers for tool access, an evaluation-and-guardrails harness, and cost/latency monitoring. This is an integration/orchestration skill that sequences several existing tool-specific skills in the correct order and flags the handoff points between them — it does not restate their internals. Use when a user asks to "build a production AI agent stack using managed LLM APIs from scratch," "stand up an agent platform with a managed vector database and MCP tools," "give me the end-to-end sequence for a cloud-managed agent deployment," or "design the full pipeline from cloud account to a production agent with evals and cost monitoring." license: Apache-2.0 compatibility: "Claude Code, GitHub Copilot, OpenAI Codex, Cursor, Gemini CLI" metadata: domain: ai-agent maturity: stable
Complete AI Agent Stack Deployment (Cloud-Managed) From Scratch
Purpose
A production AI agent is not one component — it's a chain of dependent layers (account guardrails, an agent control loop, LLM provider access, a knowledge-retrieval layer, tool access, a safety net, and cost/latency visibility) that only work as a coherent, safe system if wired up in the right order. Skip or mis-sequence a layer and the failure shows up somewhere confusing: an agent architecture designed before an LLM gateway exists hardcodes a single provider's SDK directly into agent code, making the later "add a fallback provider" phase a rewrite instead of a config change; or an MCP server is connected to a live agent before an evaluation harness exists, so the first prompt-injection-via-tool-output regression is discovered by a user, not a test suite. This skill sequences the cloud-managed version of that whole path — landing zone through cost monitoring — deliberately using vendor-neutral language for the LLM provider and vector database layers, since the managed-service choice at each of those layers is a real decision this skill defers to the reader rather than assumes.
When to use
- Standing up a new production AI agent for a team or organization that wants to build on managed cloud services (LLM provider APIs, a managed vector database, cloud landing-zone guardrails) rather than self-hosting the model-serving and vector-database layers.
- Deciding the right order to introduce an LLM gateway, RAG/vector search, MCP tool access, an evaluation harness, and cost monitoring into an agent that currently has none of them.
- Auditing an existing agent deployment for a skipped or out-of-order phase (e.g. MCP tools connected before an evaluation harness existed, or cost monitoring added only after a surprising invoice).
- Rebuilding a reference agent platform (a second product line, a second team) that should follow the same proven sequence as a known-good first deployment.
- Explaining to a team the full dependency chain for a cloud-managed agent stack build-out, phase by phase.
Prerequisites & environment
- A cloud account/subscription/project that conforms to a real landing zone — this skill does not cover account/OU, Management Group, or folder design; see whichever applies: aws-landing-zone-setup, azure-landing-zone-setup, or gcp-landing-zone-setup. Confirm outbound network egress from the workload account/subscription/ project to the chosen LLM provider(s) and managed vector database is actually permitted by the landing zone's guardrails before wiring any agent code against them.
- API credentials for at least one LLM provider (and ideally a second, for fallback), stored in a secrets manager and injected at runtime — never committed to code or the gateway's config file in plaintext.
- An account with a managed vector database service (Pinecone is the most common fully-managed choice; Weaviate/Milvus also offer managed tiers).
- A decision on which agent framework/runtime will host the control loop (a raw API loop, an agent SDK, or a CLI agent host) — this skill is framework-agnostic and assumes that choice is made independently.
- A representative set of real or realistic task inputs available before Phase 6, for building the evaluation harness — collecting these after the agent is already live is possible but produces a weaker initial eval set than gathering them deliberately up front.
Step-by-step guidance
This is the phase sequence. Each phase links to the skill that covers its full depth; the text here covers only the sequencing and integration decisions between phases.
-
Phase 1 — cloud landing zone. Confirm (or stand up) the account/ subscription/project guardrails, centralized logging, and network egress policy per whichever cloud landing-zone skill applies. Verify specifically that the workload environment's network policy allows outbound HTTPS to the LLM provider(s) and managed vector database endpoint(s) this stack will call — a landing zone's default-deny egress policy (a common, otherwise-reasonable guardrail) silently breaks every downstream phase with a generic timeout that looks like a provider outage rather than a network policy (see Common pitfalls).
-
Phase 2 — agent architecture design. Design the control loop (ReAct-style, plan-and-execute, or finite-state), termination condition, iteration cap, and tool-boundary classification (read-only vs. reversible-write vs. irreversible-write) per [agent-