When to use what

Agentic AI is a stack, not a menu

RAG, agents, MCP, and A2A are layers. Skills and rules shape behavior. Most teams need fewer layers than the pitch deck suggests; add autonomy only when the task actually branches.

RAGKnowledge — ground answers in your docs

AGENTControl loop — plan, act, observe, stop

MCPTools — standard reach into systems

A2APeers — agents coordinating across boundaries

Problem → use → example

If the real problem is…	Use	Example case
Answers ignore our docs, go stale, or invent policy	RAG	Internal policy bot cites the latest refund SOP; use GraphRAG when the question is about how entities link
Need live reads or writes against systems — not a stale index	MCP	Stock check hits ERP now; same agent opens a Jira bug with logs after a failed deploy
Only 1–2 APIs inside one app; no reuse yet	Direct tools	One Flask service calling Stripe + SendGrid; move to MCP when those same tools must serve multiple apps
Fixed path with a model in one or two slots	Workflow	Invoice PDF → extract fields → validate → post to ledger; model fills gaps, code owns order
Multi-step work that branches, retries, and must stop cleanly	Agent	Month-end expense pack: pull receipts, match policy via RAG, flag outliers, wait for Approve
Same standing orders across chats	Rules	“Never push main; always cite sources; Finnish UI copy stays formal” in every session
Named playbook with steps and a verify gate	Skill	“Author a Power BI page” skill: open model, place visuals, validate PBIR, screenshot check
Money, prod, or irreversible writes in the path	Human gate	Refund > €50 pauses for Approve; step and spend caps stop a runaway loop
Tone/format still wrong after prompts + RAG	Fine-tune	Bank chat must stick to a fixed compliance template; keep changing facts in retrieval, not weights
Specialist agents owned by different teams or vendors	A2A	Support agent hands a refund packet to Finance-ops across two tenants; each agent keeps its own MCP tools
Don’t know if the system still works after a change	Evals	20 golden tickets; block ship if pass rate drops; inspect the tool path when a case fails

Default build order: RAG → one agent with tools/MCP → stateful workflow + human gate → multi-agent inside one framework → A2A only when ownership boundaries are real.

Framework picker

LangGraphStateful, auditable, HITL, durable checkpoints

CrewAIFastest role-based prototype

LlamaIndexRAG / data-first agents

MS Agent FwAzure / .NET enterprise path

OpenAI / Claude SDKsHandoffs or Anthropic-native loops

Pydantic AISchema-first typed Python

Rules vs skills

Rules · standing orders Always on or path-scoped. Safety, voice, repo law. Keep short. Cursor: .cursor/rules, AGENTS.md.

Skills · runbooks Load when the task matches. Multi-step + verify. Cursor: SKILL.md. Don’t wrap one API call as a skill.

Memory check Rules = standing orders. Skills = runbooks. MCP = hands. RAG = library.

Ladder + gates

Prompted chat
RAG assistant with citations
Single agent + tools / MCP
Stateful workflow + human Approve
Multi-agent inside one framework
Cross-org A2A (earn this)

Kill switch Max steps, max €/tokens, tool allowlist. If you can’t stop it, you don’t ship it.

Anti-patterns

Agent for pure Q&A (start with RAG)
Treating MCP and RAG as either/or
Fine-tuning for facts that change weekly
Premature multi-agent / A2A swarms
Calling a fixed pipeline an “agent”
Unbounded autonomy where money or prod moves
Tools without schemas, budgets, or kill switches
Always-on rules that belong in a skill