Salvato in:
| Autori principali: | , , |
|---|---|
| Natura: | Recurso digital |
| Lingua: | inglese |
| Pubblicazione: |
Zenodo
2026
|
| Soggetti: | |
| Accesso online: | https://doi.org/10.5281/zenodo.19536209 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
Sommario:
- <p><strong>Episode summary:</strong> Building production agentic AI isn't about knowing one framework — it's about mastering a constellation of interconnected skills. This episode breaks down the essential technical foundations: which programming languages matter and why (Python for models, TypeScript for products), the framework landscape (LangGraph, CrewAI, AutoGen, LlamaIndex, and Claude Agent SDK), the protocols enabling agent collaboration (MCP and A2A), and the core architectural concepts (ReAct, memory systems, tool calling, and reasoning patterns) that power every serious agentic system. Whether you're prototyping or deploying to production, this is the technical map practitioners actually use.</p> <h3>Show Notes</h3> <p># The Technical Foundations of Production Agentic AI</p> <p>Building agentic AI systems that work in production requires mastery across multiple layers: programming languages, frameworks, protocols, and architectural patterns. Here's what actually matters.</p> <p>## Programming Languages: Python and TypeScript</p> <p>**Python remains non-negotiable for anything serious.** Every major agentic framework — LangGraph, CrewAI, AutoGen, LlamaIndex — is Python-first. The ML ecosystem underneath (PyTorch, Hugging Face Transformers, scikit-learn) has no peer in other languages.</p> <p>But "knowing Python" isn't enough. Agentic systems specifically demand: - **Async programming with asyncio** — agents spawn parallel tasks and make simultaneous API calls. Without async, latency compounds badly across multi-step workflows. - **FastAPI** — for building tool-serving APIs and MCP servers. - **Pydantic** — for structured tool schemas and output validation. - **Type hints** — critical for maintainability in complex systems.</p> <p>**TypeScript is increasingly pragmatic for production AI products.** It overtook Python in GitHub's 2025 language report overall. The Vercel AI SDK provides a unified interface for OpenAI, Anthropic, and Google with streaming and tool calling built in. LangGraph and the Claude Agent SDK both support TypeScript.</p> <p>The honest professional framing: Python dominates ML training and research. TypeScript leads in deploying AI to web applications. Many production systems use Python for training and TypeScript for deployment. If you learn only one, learn Python. If you're building full-stack AI products, you need both.</p> <p>## The Framework Landscape</p> <p>The framework you choose has real consequences for production systems. The landscape shifted significantly in 2024-2025.</p> <p>### LangGraph: State Machines for Agents</p> <p>LangGraph models agent workflows as directed graphs — nodes are processing functions, edges define state transitions. This handles cycles naturally, making it fundamentally better than linear chains for non-trivial tasks.</p> <p>Production users include Klarna, Cisco, and Vizient. It delivers 40-50% LLM call savings through stateful patterns, has built-in persistence with checkpointing, and supports streaming and human-in-the-loop workflows. It reached version 1.0 in late 2025 and is now the default for all LangChain agents.</p> <p>The weakness: the state graph mental model takes real time to internalize, and documentation changes frequently enough that tutorials from three months ago may not work.</p> <p>This points to a deeper ecosystem risk: 70% of regulated enterprises rebuild their agent stack every three months, according to a Cleanlab survey of 1,800+ engineering leaders. The practical implication is to keep core logic portable — prompts, tools, and evaluation harnesses should not be tightly coupled to framework-specific patterns.</p> <p>### CrewAI: Multi-Agent Teams</p> <p>CrewAI models agents as a team of specialists with roles, goals, and backstories. You define agents ("Senior Research Analyst"), define their tasks, and let the framework handle coordination. The fastest documented prototype is two to four hours from setup to working multi-agent demo.</p> <p>Enterprise users include IBM, PwC, and Gelato. It has over 100,000 certified developers in its community.</p> <p>The cost: a crew of four agents can use 3-5x more tokens than a single agent. There are documented delays on the enterprise platform (20-minute pending run delays in some configurations). For complex conditional logic, it gives less control than LangGraph.</p> <p>CrewAI has two production modes: Crews (autonomous teams with true agency) and Flows (event-driven pipelines for predictability). Flows are the more mature production pattern.</p> <p>### AutoGen: Code Execution and Dialogue</p> <p>AutoGen treats multi-agent work as structured dialogue — agents participate in group chats with defined speaking orders. The standout capability is code execution: agents write Python, execute it in a Docker sandbox, observe results, and iterate.</p> <p>Microsoft built it but shifted strategic focus to the broader Microsoft Agent Framework, merging AutoGen with Semantic Kernel. AutoGen is now in maintenance mode — bug fixes and security patches, no major features. The 0.4 release introduced async event-driven architecture and a no-code visual interface, but broke the community with incompatible changes.</p> <p>If you're in a Microsoft-heavy enterprise, it still makes sense. Otherwise, LangGraph or CrewAI are safer long-term bets.</p> <p>### LlamaIndex: RAG Specialist</p> <p>LlamaIndex occupies a different niche entirely. It's the RAG specialist — advanced indexing strategies, extensive data connectors, and superior performance for retrieval-heavy use cases. If you're building document Q&A systems, knowledge bases, semantic search, or agents reasoning over large document collections, LlamaIndex is the right tool.</p> <p>It's not competing with LangGraph on orchestration — it's the best at what it does.</p> <p>### Claude Agent SDK: Native Sandboxing</p> <p>The newest serious entrant packages the infrastructure powering Claude Code as Python and TypeScript libraries. Agents can read and edit files, run shell commands, search the web, and call external tools through MCP servers, all in a sandboxed environment.</p> <p>Setup is fast — install, provide an API key, and you're running. The weakness is model lock-in: it only works with Claude.</p> <p>## Protocols: MCP and A2A</p> <p>**MCP (Model Context Protocol)** was developed by Anthropic and is now governed by the Linux Foundation's Agentic AI Foundation, with backing from Anthropic, OpenAI, Google, Microsoft, AWS, Block, Cloudflare, and Bloomberg. It connects a single agent to external tools, APIs, and data sources. The flow: user asks, agent determines it needs external information, MCP server checks permissions, returns the result, agent responds. It's vertical integration — extending what a single agent can do.</p> <p>**A2A (Agent-to-Agent protocol)** was launched by Google with 50+ technology partners. It enables agent-to-agent communication. Agents publish "Agent Cards" — JSON self-descriptions of their capabilities — so other agents can discover and hire them. It supports parallel task execution, progress sharing, and dynamic collaboration.</p> <p>MCP and A2A are complementary, not competitive. Each agent in an A2A network might use MCP to call its own tools. MCP extends what a single agent can do; A2A expands how agents collaborate.</p> <p>There's also ACP from IBM and AG-UI for agent-to-UI communication. Three protocols racing to become the HTTP of AI agents. MCP and A2A are the serious contenders, designed to coexist. ACP may matter significantly in regulated industries.</p> <p>## Core Architectural Concepts</p> <p>Every agentic system is built on four components:</p> <p>1. **The reasoning engine** — the LLM that processes inputs, makes decisions, and plans multi-step actions. 2. **Tool calling** — how agents interact with external systems. MCP has emerged as the standard interface. 3. **Memory systems** — split into short-term working memory for the current session and long-term persistent memory stored in vector databases for cross-session continuity. 4. **Orchestration and planning** — where frameworks differ most.</p> <p>### ReAct: The Foundational Pattern</p> <p>ReAct (Reasoning plus Acting) is the foundational pattern. The agent reasons about what action to take, takes that action, observes the result, reasons about what to do next, and repeats until complete.</p> <p>Stanford's Human-Centered AI Group found that nearly 70% of multi-step tasks fail when planning mechanisms are missing. That number should alarm anyone shipping agentic systems without properly implementing the reasoning loop.</p> <p>### Advanced Reasoning Architectures</p> <p>**Chain-of-Thought** is well-known — the agent articulates its reasoning step-by-step.</p> <p>**Tree-of-Thought** explores multiple reasoning paths simultaneously before choosing one. It's computationally more expensive but significantly more robust for tasks with genuine ambiguity.</p> <p>**Reflexion** is architecturally interesting — agents reflect on their failures, learn from them, and adjust their approach on subsequent attempts. This is closer to how human problem-solving actually works.</p> <p>## The Takeaway</p> <p>Production agentic AI demands breadth across programming languages, frameworks, protocols, and reasoning patterns. The landscape is still consolidating — expect more churn in the next 12-18 months. The safest approach is to keep core logic portable and stay aware that the framework you choose today may need to be swapped in six months.</p> <p>Listen online: <a href="https://myweirdprompts.com/episode/agentic-ai-technical-foundations">https://myweirdprompts.com/episode/agentic-ai-technical-foundations</a></p>