Which tools and platforms do experienced builders use to build AI agents?

Experienced builders of AI agents use a multi-layered stack that spans orchestration frameworks, specialized infrastructure, and robust context management systems. The focus has shifted from simple "model wrapping" to building complex "harnesses" that allow models to use tools in a recursive loop.

Orchestration and Frameworks

Orchestration is the "brain" layer that manages how an agent reasons and invokes tools.
* LangChain and LangGraph: These are widely used to build Directed Acyclic Graphs (DAGs) for agents. LangGraph, in particular, is noted for its ability to handle complex agentic loops and state management Y Combinator — Now Anyone Can Code @ 12:16.
* CrewAI: This tool is frequently used for multi-agent orchestration, allowing builders to create "crews" of specialized agents. It includes features for local testing and quality scoring Greg Isenberg — How To Create Ai Agents From Scratch @ 09:14.
* DSPY: This framework is used for optimizing system prompts by treating them as programs that can be "learned" and iterated upon through genetic programming Y Combinator — Why The Harness Matters More Than The Model @ 12:16.
* AutoGPT: Cited as an early pioneer in the space, it is often used by builders looking to experiment with autonomous, real-time decision-making Greg Isenberg — The Secret Strategies of Top DTC Brands with Nik Sharma @ 27:23.

Infrastructure and Platforms

Deploying agents requires infrastructure that can handle long-running tasks and "untrusted" code execution.
* Vercel: Builders use Vercel’s AI Gateway and sandboxes to run code generated by agents securely. It also offers "fluid compute" for efficient resource allocation during complex workflows Lenny's Podcast — What world-class GTM looks like in 2026 @ 42:44.
* Replit: Replit provides a "multiplier editor" and specialized AI Computer Interfaces (ACI) that allow agents to interact directly with shells, package managers, and file systems rather than relying purely on vision-based interfaces Lenny's Podcast — Behind the product: Replit @ 30:45.
* MCP (Model Context Protocol): This protocol is gaining traction as a standard way to give agents (like Claude) access to external tools and data sources without custom, one-off integrations Greg Isenberg — I built next gen AI Agents with MCP + Claude @ 03:06.

Context and Memory Management

Advanced builders are moving beyond basic Retrieval-Augmented Generation (RAG) toward more structured memory systems.
* Convex and Postgres: Convex is favored for real-time applications like chat agents because it allows builders to define and interact with the database directly in code a16z — Why Creativity Will Matter More Than Code @ 57:52. Postgres is often used to centralize agent traces and session data for human review Y Combinator — Why The Harness Matters More Than The Model @ 51:58.
* Obsidian: For solo builders, maintaining a structured wiki of markdown files in Obsidian provides high-quality context for agents to understand long-term projects and personal preferences Greg Isenberg — The $1M+ Solo AI Agent Business @ 24:31.

Preferred Models

While builders often use multi-model architectures, Claude 3.5 Sonnet is currently cited as the top choice for coding-specific agents due to its superior performance in generating and editing code Lenny's Podcast — Behind the product: Replit @ 30:45 Y Combinator — Now Anyone Can Code @ 03:03.

— Sources: 12 videos across 4 creators