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

Experienced builders utilize a combination of open-source orchestration frameworks, specialized coding environments, and standardized protocols to build and deploy AI agents. The consensus among creators like Y Combinator and Greg Isenberg is that building agents is moving away from simple "prompt-in, response-out" patterns toward autonomous loops that "Observe, Think, and Act" Greg Isenberg — Building AI Agents that actually work (Full Course) @ 01:42.

1. Orchestration Frameworks

Builders use frameworks to manage "multi-agent orchestration," where different models handle specialized sub-tasks.
* CrewAI: Frequently highlighted for creating "crews" of agents that can be tested locally using a CLI (crewai create). It allows builders to assign "senior" roles to agents to steer their behavior and provides "task scores" to evaluate quality Greg Isenberg — How To Create Ai Agents From Scratch (CrewAI, Zapier, Cursor) @ 05:54.
* LangChain and Deep Agent: Mentioned by Jensen Huang as essential tools for companies building domain-specific AIs Y Combinator — Jensen Huang: The Mindset That Built NVIDIA @ 27:27.
* OpenCL / OpenClaw: Used by Garry Tan and others at Y Combinator to build "personal AIs" that diarize and manage massive amounts of personal data across thousands of markdown files Y Combinator — Garry Tan: Own Your Intelligence @ 12:16.

2. Specialized Agent Coding Environments

For technical builders, the choice of environment often dictates the agent's effectiveness:
* Replit Agent: Amjad Masad explains that Replit provides the underlying infrastructure (editor, package managers, shell access) directly to the AI, allowing it to build, test, and deploy applications autonomously Lenny's Podcast — Behind the product: Replit | Amjad Masad (co-founder and CEO) @ 30:45.
* Cursor and Windsurf: These are widely adopted for "vibe coding," where instructions are stored in files like .cursorrules to provide the agent with deep context Y Combinator — How To Get The Most Out Of Vibe Coding | Startup School @ 09:13.
* Claude Code and Codeex: Described as "agent harnesses" that facilitate the agent loop specifically for development tasks Greg Isenberg — Building AI Agents that actually work (Full Course) @ 04:14.

3. Integration and Tooling Protocols

To give agents "hands," builders rely on connection protocols:
* Model Context Protocol (MCP): Developed by Anthropic, this acts as a standardized "translator" allowing agents to interact with external tools like Notion or Slack without custom code for every integration Greg Isenberg — Building AI Agents that actually work (Full Course) @ 18:45. Bret Taylor notes that advanced teams even dedicate engineers to maintaining their own MCP servers to ensure agents have the right context Lenny's Podcast — He saved OpenAI, invented the “Like” button, and built Google Maps: Bret Taylor (Sierra) @ 1:09:05.
* No-Code/Low-Code: For faster deployment, builders use Lindy.AI for "AI employees" or n8n and Zapier to bridge gaps between models and existing business apps Greg Isenberg — OpenAI's NEW Agent Builder and ChatKit are INSANE @ 00:00.

4. Hardware and Infrastructure

Y Combinator suggests that experienced builders are starting to look beyond standard GPUs. Current chips are inefficient for the "bursty" nature of agent loops (backtracking and tool calling), leading to a rise in interest for purpose-built silicon like Groq that supports fast context switching and persistent KV caches Y Combinator — Inference Chips for Agent Workflows @ 00:00.

— Sources: 12 videos across 4 creators