Agentic AI Patterns and Anti-Patterns

AI Agents are the buzzword of the year. Yet relying on the magic of autonomy remains risky due to hallucinations and planning errors.

After a quick overview of the agent personas, and defining the agent equation, we'll go through architectural patterns like the conductor for task supervision, rethinking tools for lower hallucination rates, and using proven and promising protocols like Model Context Protocol (MCP) for standardizing tools and Agent 2 Agent protocol (A2A) to foster an ecosystem of agents.

We will also dismantle common anti-patterns, such as the leadership's chatbot mandate instead of augmented UIs, the silent confabulation where users can't trust the agent responses, and showing you how to build systems that focus on verifiable results and real business value.

Transcript

1. Agentic AI patterns & anti-patterns Guillaume Laforge Developer Advocate Google @glaforge
2. Unavoidabl e wave 1
3. From Transformers to Autonomous Agents
4. 2017 / Google Research invents Transformers
5. 2022 / OpenAI popularizes LLMs with ChatGPT
6. 2024 / Retrieval Augmented Generation trending
7. 2025 / AI Agents, Buzzword of the Year
8. The Three Agent Personas
9. The Three Agent Personas
10. The Three Agent Personas
11. The Three Agent Personas
12. The Agent Equation
13. The Loop : Observe > Plan > Act THINK Analyze the prompt, system instructions to define the goal to reach PLAN Check available tools, define the strategy to reach the goal ACT RAG search, API call, execute code, invoke other agents, Human-in-the-Loop REFLECTION Think & evaluate the output to fix mistakes, suggest improvements and finalize the result
14. Architectural Patterns 2
15. The Magic of Autonomy? Even the best LLMs still hallucinate with “function calling” • Non-existent function • Wrong invocation order • Hallucinated parameters The system giga prompt confuses LLMs (cf “context engineering”) Autonomous agents ⇔ Workflows (“agency” vs determinism)
16. A “conductor” supervises sub-agents Breaking down a complex task into more focused sub-tasks Possible approaches: • Graph-oriented frameworks (LangGraph) • Frameworks with a hierarchy of agents (ADK, LangChain4j…) • Workflow tools (n8n) • Programmatic Pattern #1 / The Conductor
17. ⚠ It’s not enough to expose a REST API as tools to an LLM Think about the concrete business tasks the agent needs to perform (ex. scheduling a meeting) Fewer tools ➜ less confusion ➜ more determinism Select, filter, group the tools Delegate a subset of tools to a sub-agent Pattern #2 / Rethinking Tools
18. Model Context Protocol invented by Anthropic The “USB” of outils : a protocol for the interaction between an agent and its tools No more need to implement the integration glue in each project ➜ server configuration Protocol widely adopted today, but still evolving The S in MCP stands for Security… Pattern #3 / MCP, Standardizing Tools
19. Agent 2 Agent Protocol launched by Google and multiple partner, and now at Linux Foundation Discovery of other agents through an agent identity card: description of skills Multimodal exchange of messages, tasks, and artifacts ➜ For interoperability between agents of all languages or frameworks Pattern #4 / A2A, Towards an Agent Ecosystem
20. Frontier and Anti-Patterns 3
21. The mandate from the leadership to add a chatbot to the product User interactions do not need to be conversational If there’s a conversation, think multimodal and rich UI ➜ The best AI agent integration should be transparent like a Head-Up Display (HUD) Anti-Pattern #1 / The Chatbot Mandate
22. Vibe-checking the result of your Agent is insufficient Working hand-in-hand with subject-matter experts Collecting real questions and interactions (golden responses) Employing techniques like the RAG triad, LLM-as-Judge (with tools like DeepEval, PromptFoo, RAGAS…) ➜ Evaluation phase is essential Anti-Pattern #2 / Insufficient Vibe-Checking
23. Retrieval Augmented Generation and “search grounding” : ➜ Cite your sources! (links, pop-over, overlay…) Give confidence in generations LLms invent responses with great force and conviction! (“You’re absolutely right!”) ➜ IVO approach to visually validating answers Anti-Pattern #3 / Silent Confabulation
24. The augmented developer Coding agents can lead you down the wrong path With assurance, conviction, and self-confidence! Look up and focus on added value ➜ MVP vs “feature creep” Anti-Pattern #4 / Down the Rabbit Hole
25. Lundi Matin… Back at the Office 4
26. Don’t ask yourself where to add a chatbot! Identify the CUJ, the most painful business process that can be improved Experiment, launch your first agent, because the goal is to learn Measure: Your users will validate the relevance of your choices Can agents buy happiness? :-) 1 2 3 4
27. Thanks for your Attention (is all you need) @glaforge