// field guide
The agent shipyard: from prototype to production
A drydock is where ships are built, repaired and made seaworthy before they launch. Building AI agents feels much the same. A demo that works once in a notebook is a long way from an agent that runs reliably for thousands of users, and the gap is filled by infrastructure: frameworks that structure the reasoning loop, orchestration that survives failures, memory and retrieval, fast inference, sandboxes for running code, and protocols that connect agents to the tools they need. Drydock is a directory of that infrastructure, written for engineers who ship.
Most projects start with an agent framework. LangChain and LangGraph offer the broadest integrations and fine grained graph control; LlamaIndex excels when your agent works over documents; CrewAI and AutoGen focus on teams of collaborating agents; the OpenAI Agents SDK, Pydantic AI and smolagents keep things lightweight; and Mastra and the Vercel AI SDK serve TypeScript teams. If you prefer visual building, Dify, Flowise, Langflow and n8n let you wire agents together on a canvas. For long running work, durable execution engines like Temporal and Trigger.dev make sure a crash halfway through does not mean starting over.
Agents need memory and models. Vector databases (Pinecone, Weaviate, Qdrant, Chroma, Milvus, pgvector and LanceDB) store embeddings for retrieval and recall, and memory layers like Mem0 and Letta manage what an agent remembers about users across sessions. On the inference side, vLLM and SGLang serve open models at high throughput, Ollama and LM Studio run them locally, LiteLLM and OpenRouter put many providers behind one API, and platforms like Together, Fireworks, Groq, Modal, Replicate and Baseten handle scaling for you.
Coding agents are where many developers meet agents first. Claude Code, Codex, Gemini CLI, Aider and Cline operate on real repositories; Cursor, Copilot and Zed bring agents into the editor; Replit and Lovable build whole apps from a prompt. Underneath it all, the Model Context Protocol is becoming the standard way to plug tools into agents, while E2B, Browserbase, Firecrawl, Exa and Tavily provide sandboxes, browsers, web data and search. Every listing shows live GitHub stars where the project is open source, and the feed below tracks new agent repos, papers and launches.