Source-level comparisons of today's agent frameworks — what each one ships for error handling, observability, deployment, tool orchestration, human-in-the-loop, and agent-to-agent communication. Drawn from analyzing the frameworks' own source code.
A from-scratch, in-depth explanation of Agent Skills — the concept and its implementation. What a skill is, the SKILL.md format, the progressive-disclosure trick that keeps them cheap, how a skill is triggered and run, and the full implementation across the Claude Developer Platform API, Claude Code, the Agent SDK, and the open agentskills.io standard — plus authoring best practices and security. Cited to official documentation.
Read paper →Three concepts agent builders constantly conflate — tools (function calling), the Model Context Protocol, and Agent Skills — are three different layers, not three competitors. This paper draws the line: tools are the unit of action, MCP the unit of integration, skills the unit of know-how — who executes what, how they stack into one system, and when to reach for which. Cited to official documentation.
Read paper →A JSON-first reference across nine provider APIs — OpenAI (Chat Completions and Responses), Anthropic, Google Gemini, Ollama, Mistral, Cohere, AWS Bedrock, Groq, and xAI. It draws the boundary between your agent runtime and the model provider, then shows the request and response shapes each one puts on the wire — cited to official documentation.
Read paper →A code-first reference across ten AI agent frameworks — LangChain, LangGraph, CrewAI, Google ADK, Semantic Kernel, Mastra, Microsoft Agent Framework, LangChain4j, Genkit Go, and Rig — covering how each one declares tools, defines their schemas, and composes multiple agents.
Read paper →A reference map of where tools can live across ten AI agent frameworks — in-process code, MCP servers, OpenAPI specs, remote agents over A2A, and generic RPC — and the documented declaration and call mechanism for each.
Read paper →A protocol-level reference for MCP 2026-07-28 — its stateless JSON-RPC request model, stdio and POST-only Streamable HTTP transports, discovery, tools, caching, Multi Round-Trip Requests, subscriptions, optional OAuth 2.1 authorization, and deployment patterns for local AI hosts and server-hosted agents.
Read paper →A side-by-side conceptual tour of the agent harness inside LangChain, LangGraph, CrewAI, Google ADK, and Microsoft Semantic Kernel — what the loop looks like, who holds state, and how each framework decides when to stop.
Read paper →How agents in LangChain, LangGraph, CrewAI, Google ADK, and Microsoft Semantic Kernel find each other, expose their capabilities to the LLM, and talk — both embedded in the same process and distributed across services.
Read paper →How LangChain, LangGraph, CrewAI, Google ADK, and Microsoft Semantic Kernel let a person step into the agent loop — to approve a tool call, edit a plan, sign off on the final answer, or own a long-running decision the agent simply cannot make alone.
Read paper →Production agent runs fail in many more ways than the four buckets the frameworks themselves design around. Five chapters cover what actually breaks in production — in-loop failures, provider-side failures, multi-agent failures, operational failures, and data & compliance failures — and what LangChain, LangGraph, CrewAI, Google ADK, and Microsoft Semantic Kernel do (or don't do) about each one.
Start with the index →What does it take to run an agent in production — observe it, scale it, deploy it, test it, evolve it — across ten frameworks spanning Python, TypeScript, .NET, Java, Go, and Rust? Five chapters: each focused on one operational axis (observability, performance, deployment, testing, and configuration/memory) with source-grounded comparisons and practical takeaways.
Start with the index →