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.
Complex tasks often need multiple LLM calls, but deciding how to split the chain, pass intermediate results, and recover from errors is an engineering problem. This article covers sequential, parallel, routing, and loop patterns; when to use chains versus Agents; state management; and cost-quality tradeoffs.
Letting an Agent open a browser, click pages, scrape data, and fill forms—agentic browsing is a high-value AI automation scenario and a deep engineering challenge. Covers architecture (Playwright/Puppeteer + vision models), operation modes (coordinate clicks vs.
Single-turn conversations are easy; multi-turn conversations are hard. Users follow up, change their minds, refer to prior context, and interrupt. This article covers context management, dialogue state, coreference resolution, topic switching, interruption, and resumption so conversations remain coherent.
Agents are non-deterministic, multi-step, and call tools — manual testing can't keep up. This article explains how to automate agent evaluation: build task sets + rubrics, batch-score with LLM-as-judge, do trace-driven regression, wire eval into CI, and mine bad cases from production traces for auto-intake — so every agent change gets an objective score.
MCP lets AI tools plug into your capabilities, but also exposes data and operations to every compatible client. This guide covers least-privilege tools, validation and injection defenses, remote authentication and rate limiting, local-vs.-remote trust boundaries, third-party supply-chain risk, and a prelaunch checklist.
An Agent is a loop at its core, but whether that loop is well-designed decides if it is intelligent or just thrashing. This article covers the key design points of the Agent loop: stop conditions (when to end), self-reflection (self-assess after each step), branching and backtracking, max steps and budgets, and adding "human checkpoints" to intercept
Attack before launch, or users will afterward. LLM red teaming finds prompt injection, jailbreaks, privilege abuse, data exfiltration, harmful output, and denial of service. This article covers adversarial prompt sets, LLM-generated attacks, and turning findings into regression tests so you discover flaws before users do.
With multiple Agents, the hardest problem is not individual capability but communication: shared state or message passing, protocols, asynchrony, and errors. This article explains blackboards, message passing, handoffs, serialization and contracts, supervisor/pub-sub/peer-to-peer architectures, plus engineering concerns around serialization, versioning, and