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.
Parallel tool branches completing within the same plan does not mean the fastest result may modify state first. This article uses branch-result staging, dependency closure, decision versions, and a transactional commit barrier so out-of-order results produce only one explainable downstream advance.
An Agent without memory starts from zero every time—it can't truly collaborate continuously. This article explains the three-tier memory model (working/short-term/long-term), storage and retrieval of long-term memory, the most easily overlooked write strategy and forgetting mechanism, and how to build a production-grade Agent memory system—memory isn't better when there's more, it's better when it's accurate.
AI support is among the easiest Agent use cases to deploy, and to turn into a "dumb bot." This article builds a deployable support Agent end to end: RAG for FAQs, function calling for order lookups and ticket creation, and human handoff when uncertain.
Pre-executing tools the model may choose can reduce latency, but it can also send messages, reserve inventory, or pollute caches before the user decides. This article combines speculation tickets, isolated execution, verifiable dry runs, and atomic promotion into a strict speculative boundary.
The best way to understand an Agent is to build a minimal one yourself. This under-100-line TypeScript implementation explains the ReAct (Thought → Action → Observation) loop, tool calls, stopping conditions, and error recovery.
Use MCP to let Claude Code read issues, open PRs, inspect CI, edit files, and submit changes. This guide covers the official remote MCP server, a self-hosted local server, configuration, practical workflows, and safe token permissions so AI-assisted development can close the loop in GitHub.
Tool-call failures can originate in selection, arguments, authorization, dependencies, state, or user confirmation. This article provides an actionable taxonomy and metric design to turn vague Agent errors into fixable engineering problems.
Deep research—one question, multiple search rounds, and a synthesized report—is one of the most valuable agent use cases and cannot be solved by single-pass RAG. This article builds a LangGraph research agent that decomposes questions, searches in parallel, reads full text, fills gaps, and writes a report while explaining state graphs, tool calls, and loop
Passing the latest validator does not mean legacy Agents, renderers, and audit jobs still in production can read a new result. This article combines producer revisions, the real consumer set, decoding proofs, and representation routing into a compatibility gate purpose-built for tool results.
Testing agents is harder than ordinary software—non-deterministic, multi-step, with external tools. This article gives a layered strategy: unit tests for tools, integration for end-to-end, evaluation sets for quality, chaos for fault tolerance, plus trace reproduction and regression. Turn "launching on luck" into "launching with confidence".
AI framework choices are unavoidable, but LangChain, LangGraph, and LlamaIndex are often complementary, not competitors. This article maps their roles: full-stack glue, agent orchestration, and data/RAG; explains how to choose; and covers an underrated option: using no framework and writing it yourself.
A single Agent working doesn't mean it can serve multiple users and tasks. This article clarifies the concurrency model (per-request Agent vs Agent pool), resource management (model API concurrency limits / tool concurrency), scheduling (priority / queue / rate limiting), and multi-user state isolation, pushing the Agent from a "single-threaded demo" to a "traffic-bearing service".