AI Agent Architecture CHEAT SHEET

AI Agent Architecture CHEAT SHEET 1 LLM (Brain) The reasoning engine that understands instructions and generates responses. Instructions Context Response / Action LLM (e.g. GPT-4, Claude, Llama, etc.) ✓ Understands, reasons and generates ✓ Follows instructions and uses tools ✓ May hallucinate or make mistakes Use when you need natural language understanding, reasoning or generation. 2 Memory (Retains Context) Stores information so the agent can remember and use it later. Short-term Memory (Conversation History) Long-term Memory (User Preferences, Past Interactions) Episodic Memory (Events, Experiences) ✓ Enables continuity and personalization ✓ Improves decisions with past context ✓ Large memory can be costly Use when your agent needs to remember chat history, user info or past events. 3 RAG (Knowledge Access) Retrieves up-to-date or domain-specific knowledge to ground the agent. Query Vector DB (Docs, Data) Embeddings + Search Relevant Context ✓ Reduces hallucinations ✓ Enables private, up-to-date knowledge ✓ Depends on quality of data Use when LLM has external knowledge that the LLM should use. 4 Tools (Action Executor) Allows the agent to interact with real world or external systems. Search (Web) Calculator Database API / Service Email / Message Action Result ✓ Extends capabilities beyond LLM ✓ Makes agent practical and useful ✓ Tools may fail or return errors Use when the agent needs to perform actions or fetch real-world data. 5 Planning (Decide Steps) Breaks down complex goals into small, actionable steps. 1. Understand Goal 2. Break into Steps 3. Decide Next Action 4. Execute & Observe 5. Repeat until Done ✓ Handles complex, multi-step tasks ✓ Improves success rate ✓ Over-planning can be slow Use when the goal is complex and requires multiple steps or tool usage. 6 Reflection (Self-Improve) Evaluates results, finds mistakes, and improves future decisions. Observe Output Evaluate Identify Issues Improve & Retry ✓ Increases accuracy over time ✓ Helps recover from mistakes ✓ May increase token usage Use when: Accuracy is important and the task looks for improvement. 7 Multi-Agent Collaboration Monitors agents working together, each with a specific role. Manager Agent (Orchestrator) Research Agent QA Agent Coder Agent Shared Memory / Knowledge Base ✓ Parallel work = faster results ✓ Fewer agents = harder to design and debug 8 Monitoring & Guardrails Ensures the agent is safe, reliable and aligned with your rules. Input Validation Output Filtering Safety Checks Rate Limiting Logging & Audit ✓ Prevents harmful or unsafe outputs ✓ Ensures reliability and safety ✓ Too strict rules can block good outputs 9 Human-in-the-Loop Reviews, guides or approve important decisions. Human Output / Recommendation Approve / Edit / Reject Continue Execution ✓ Improves accuracy and trust ✓ Essential for hard or repetitive tasks ✓ Slower and requires human effort Key Takeaways ✓ Similarity search finds the most relevant results ✓ Indexes enable fast search at scale ✓ Essential for RAG, semantic search, recommendations, and more ✓ Indexes enable fast search at scale ✓ Metadata and filtering improve retrieval quality