#智能体架构
32 related articles
LangChain Breaks Down Self-Improving Agents: The Triple Evolution of Models, Harnesses, and Context
LangChain breaks down the three dimensions of self-improving agents: models (SFT and RL), harness, and context—and how traces, evals, and red teaming form the flywheel.
New Trends in LLM Engineer Interviews: Why Simple Agents No Longer Land You an Offer
LLM interviews are leveling up—simple agents no longer suffice. This article breaks down task planning, multi-model backups, and tool orchestration architecture in enterprise-grade complex agents, helping engineers understand the system design skills interviewers really assess.
Stop Strangling AI Agents with Workflows — Give Them "Childproofing" Instead
A Reddit developer proposes "childproofing" over rigid workflows: hard environment boundaries, code verification, and test-driven evals let models self-plan.
Stop Using LangGraph to Orchestrate AI Agents — Refactor Your Reasoning Harness with 'Child-Proofing'
In 2026, AI agent development is stuck in rigid LangGraph-style workflows. This article presents a 'child-proof' reasoning harness: replace flowchart orchestration with hard environment boundaries, external code validation, and test-driven evaluation — letting models plan and execute autonomously while staying safe and controllable. Includes a Customer Pulse prototype and a candid trade-off analysis.
How Anthropic Decoupled the Brain and Hands of Claude Agents
At AI Engineer, Anthropic reviewed Claude Managed Agents' evolution: from a single-container "pet" architecture to a session log, harness brain, and sandbox hands fully decoupled, cutting first-token latency ~90% and keeping credentials out of the sandbox to resist prompt injection.
How Can Frontend Developers Advance into AI Agent Development? Deconstructing System-Level Agent Architecture from OpenClaw
Starting from the OpenClaw demo in a Bilibili open class, this article breaks down the three-layer architecture and universal formula of system-level AI agents, and analyzes the path for frontend developers to enter agent development based on TypeScript, including plugin-based design and security/memory layer challenges, with an interview answer framework.
AI Agents vs. ChatGPT: The Paradigm Shift from Passive Responses to Autonomous Action
Explaining the essential differences between AI Agents and ChatGPT: from passive chatbots and preset AI workflows to LLM-driven autonomous agents, covering the ReAct framework, self-correction loops, and human-in-the-loop safeguards.
From Human-in-the-Loop to Out-of-the-Loop: A Hands-On Guide to Automated News Curation with Durable Agents
IndieDevDan shows how Durable News Agent uses durable agents and triggers to curate news automatically, filtering hype, facts, and signals with feedback weights.
Accelerating Hermes Agents with Jev: An Engineering Practice of Fast-Slow Model Division
Explains how to accelerate the Hermes agent framework with the lightweight Jev model: by dividing work between fast and slow models, simple intent judgment response time is compressed from several seconds to about 200 milliseconds, with analysis of the applicable scenarios and boundaries of layered decision architecture.
Inside Redis's Playbook: Context Engineering for Production-Grade AI Agents
Redis's AI team shares real-world context engineering for production AI agents: context window management, tool abstraction, security, evals, and traceability.
AML Agent in Practice: Optimizing Memory Is More Effective Than Swapping LLMs
How to build an AML agent? Based on AI developer Nacho Martinez's tech talk, this article analyzes the memory layer, context engineering, and tool scheduling in agent architecture, explaining why optimizing the scaffolding beats swapping LLMs for financial risk control tasks.
Why 86% of Multi-Agent Systems Fail: Unpacking Planner-to-Worker Spec Drift
Production multi-agent systems fail at rates up to 86.7%, driven by planner-to-worker spec drift and silent tool error propagation. This article breaks down three handoff bottlenecks, using an audit of 10,000 workflows to show how deterministic state contract graphs can raise end-to-end success to 96.9%.
Harness Architecture in Practice: A Full Breakdown of Enterprise-Grade Agents from Deployment to Production
A hands-on enterprise agent project on the Harness architecture: Vue + FastAPI, ASGI streaming, MCP gateway for ERP, sandbox isolation for Skill security and user data.
Production-Grade Multi-Agent Systems: When AIs Contradict Each Other
A developer's six-month production retrospective on multi-agent systems: orchestration pattern choices, circuit breaker design, idempotency implementation, and observability—revealing distributed state contradictions and graceful failure strategies among AI agents.
MCP vs Function Calling: A Decision Framework for Agent Tool Selection
MCP vs LLM function calling: compare core differences, MCP's hidden operational costs, and a practical decision framework for agent tool selection.
Why AI Agents Need Their Own Sandbox: A Deep Dive from a Google DeepMind Engineer
Google DeepMind engineer Philip explains Gemini API's Interactions API, remote sandbox, and Agents API at AI Engineer, and why every AI agent needs its own sandbox.
OpenSeal Agent Architecture Deep Dive: Persistent Orchestration and Approval Governance
A deep dive into OpenSeal's AI agent architecture: persistent execution via Objective/Run/Turn/Skill, lease recovery, capability checks, approval governance, and model routing design.
Inside Always-On AI Agents: Persistence, Proactivity, and Long-Running Task Architecture
Stanford CS146S: A deep dive into always-on AI agent architecture — from four developer eras and six model capabilities to thin-client/thick-server design, subagents, prompt caching, and compaction.
How Uber Solved MCP's Biggest Context Problem with Omni MCP
Uber's Omni MCP proxy server solves context explosion at 5,000+ tools using four meta-tools: discover server, discover tools, get tool schema, and invoke tool. A deep dive into its architecture, permissions, and response projection.
The Hidden Token Tax of MCP: One Server Burns 18,000 Tokens Before Doing Any Real Work
A single Chrome DevTools MCP server can burn 18,000 tokens before doing any work. Learn how to measure, prune, lazy-load, and optimize your MCP tool catalog.