Agentic AI & multi-agent systems templates
Single agents, multi-agent orchestration, MCP and A2A protocols, RAG variants, memory, guardrails and LLMOps — the architectures teams are building AI products on. Every one opens as a real, editable diagram — restyle it, extend it and export it as your own. See also the system architecture diagram tool.
All 21 agentic ai & multi-agent systems templates
ReAct agent (reason + act loop)
The basic agent loop: the model reasons, picks a tool, observes the result and repeats until it can answer — the pattern under most agent frameworks.
MermaidSupervisor multi-agent orchestration
A supervisor agent routes work to specialist agents (research, coding, data, writing), collects results and decides when the task is done.
MermaidHierarchical agent teams
A top-level manager delegating to team leads, each running its own workers — how large tasks like building a product brief or a codebase are decomposed.
MermaidAgent handoffs (swarm pattern)
Agents pass the conversation to each other with context — triage to billing to technical support — without a central supervisor.
MermaidOrchestrator-workers pattern
An orchestrator LLM breaks a task into subtasks at runtime, fans them out to worker LLM calls in parallel, and synthesises the results.
MermaidPrompt chaining with gates
A fixed workflow of LLM steps with programmatic checks between them — outline, check, draft, translate — cheaper and more predictable than a free-running agent.
MermaidLLM gateway & model routing
A gateway in front of several model providers that routes by task and cost, caches responses, enforces limits and falls back when a provider fails.
MermaidAgent-to-agent (A2A) network
Independent agents from different teams or vendors discovering each other via agent cards and delegating tasks over the A2A protocol, with MCP for their tools.
MermaidModel Context Protocol (MCP) architecture
An AI host app with MCP clients connecting to local and remote MCP servers that expose tools, resources and prompts.
MermaidAgentic RAG with self-correction
A retrieval agent that rewrites queries, grades retrieved documents, falls back to web search and checks its answer for hallucinations before responding.
MermaidGraphRAG with a knowledge graph
Documents turned into an entity-relationship knowledge graph plus community summaries, then queried together with vector search for multi-hop questions.
MermaidDeep research agent
A long-running agent that plans research, searches and reads in parallel, reflects on gaps, and writes a cited report.
MermaidAutonomous coding agent
How a coding agent works on a ticket: explore the repo, plan, edit, run tests, fix failures and open a pull request for human review.
MermaidComputer-use / browser agent
An agent that operates a real browser or desktop by looking at screenshots and issuing clicks and keystrokes, with safety confirmations.
MermaidAgent memory architecture
Short-term context, episodic and semantic long-term memory, and procedural memory — how agents remember across sessions.
MermaidLLM guardrails pipeline
Input and output guardrails around a model call: prompt-injection and PII checks going in, grounding, toxicity and policy checks coming out.
MermaidLLMOps: evaluation & monitoring loop
Offline eval sets, CI gates on prompt changes, online tracing, user feedback and regression datasets — how teams ship LLM features safely.
MermaidMulti-agent debate & critique
Several agents propose answers, critique each other over rounds, and a judge picks or merges the result — used to improve reasoning accuracy.
MermaidLLM fine-tuning pipeline
Collect and clean data, create instruction pairs, fine-tune with LoRA, evaluate against the base model and deploy behind an A/B test.
MermaidAI customer service agent with escalation
An agent answering support with the knowledge base and account tools, taking safe actions itself and escalating to humans with a summary.
MermaidEnterprise agent platform architecture
The platform layer companies build for many agents: registry, tool catalogue via MCP, identity, policy, memory, observability and a model gateway.
Mermaid