AI Agent、MCP、工具调用、多 Agent 协作
This collection features articles and community discussions in the Agent 工程 space, focusing on real-world engineering scenarios, tool selection, practical paths, and risk boundaries.
The page is continuously updated and serves as a learning roadmap, technology selection reference, and technical review entry point.
Model Context Protocol (MCP) is Anthropic's "USB-C interface for AI applications." This article explains the host/client/server model, resources/tools/prompts, and JSON-RPC, then builds a TypeScript MCP Server for Claude Code from scratch.
"Agent ≠ multi-turn dialogue." Breaks down the four components of an AI Agent—planning, memory, tools, execution—and explains ReAct vs. Plan-Execute, short/long-term memory design, function calling engineering points, and pushing an Agent from demo to production via loop control, error recovery, and observability.
After a user confirms a high-risk action once, repeated clicks, queue re-sends, and parameter substitution can all reuse the approval. This article closes those windows with a canonical action digest, one-time consumption, state versions, and idempotent execution.
The most common problem in multi-agent systems isn't model capability, but incomplete inputs, unclear responsibility, and unverifiable results after task delegation. This article covers delegation contracts, result schemas, budget propagation, and recovery mechanisms.
Letting agents freely fetch URLs turns the model into an intranet request proxy. This article covers URL normalization, full-address validation, connection IP pinning, per-hop redirect inspection, egress proxies, and an attack regression suite.
The previous article covered the MCP protocol; this one is hands-on. Build a production-ready MCP Server from scratch: pick a scenario, use the official SDK to implement resources and tools, deploy locally via stdio or remotely via HTTP, debug with MCP Inspector, then wire it into Claude Code and Cursor.
An agent's evaluation set determines what the team will optimize for. This article covers how to sample from real workflows, annotate observable outcomes, cover failure paths, and maintain versions, so evaluation results actually guide engineering decisions.
Technical trends, engineering practices, and an independent developer perspective on MCP and AI Agents.
Technical trends, engineering practices, and independent developer perspectives around AI Agents and large language models.
An independent developer's perspective on technical trends and engineering practices involving RAG and AI agents.
When an Agent errors, it should not mechanically retry. This article provides a method to decide the recovery path based on error retryability, idempotency, risk level, and evidence completeness, and explains when to degrade or hand off to a human.
Technical trends, engineering practices, and independent developer perspectives around AI Agents and AI.