BestBlogs Issue #81: Long-Running Agents

This article presents a comprehensive framework for writing high-quality specifications (Specs) for AI coding agents, moving beyond "vibe coding" toward rigorous AI-assisted engineering. Key principles include starting with a high-level vision, utilizing a professional PRD-like structure (covering commands, testing, and style), and modularizing tasks to mitigate the "curse of instructions." It introduces the "Always/Ask First/Never" boundary system to maintain control over AI actions and highlights the importance of Plan Mode and the Model Context Protocol (MCP). This is an essential read for developers seeking to build reliable, maintainable systems in collaboration with increasingly powerful AI agents.

In this insightful episode, LangChain co-founder Harrison Chase explores the paradigm shift from simple LLM scaffolding to sophisticated "Agent Harnesses." He argues that in the realm of Long Horizon Agents, the traditional "code is truth" mantra is being replaced by "Traces" as the definitive source of truth due to the non-deterministic nature of AI. The discussion delves into the intricacies of Context Engineering, the critical role of file systems, and the rise of "First Draft" applications in coding and research.

This article provides a profound analysis of the two pillars in the AI Agent ecosystem: MCP (Model Context Protocol) and Skills. By using the analogy of MCP as a "USB protocol" and Skills as "applications," the author clarifies their distinct roles. A key highlight is the critical evaluation of MCP's overhead, where tool definitions can consume up to 40% of the context window, leading to high costs and reduced accuracy. In contrast, it introduces Skills' "Progressive Disclosure" and script-based execution as a more token-efficient alternative for local workflows.

This article provides an in-depth guide to building a professional "Prompt Optimizer" using Claude Skills and AI coding tools. Addressing common friction points in prompt engineering—such as vague requirements and lack of structured frameworks—the author introduces an automated solution leveraging over 50 global prompt frameworks. It details a complete workflow from defining Skill specifications to developing a Chrome extension. The key takeaway is that in the AI era, the cost of "doing" has plummeted, making "thinking" and task decomposition the primary value-adds.

This article provides a profound analysis of the paradigm shift in software development—from manual coding to "intent-driven" engineering. It systematically explores the evolution from intuition-based Vibe Coding to the disciplined Vibe Engineering framework. By revisiting classical theories of Essential vs. Accidental Complexity, the author argues that while AI minimizes technical implementation hurdles, the core business logic remains a challenge that requires Domain Modeling and Spec-Driven Development (SDD).

This article provides a deep dive into the engineering practices of Cursor and Anthropic for building "long-running" AI agents. While Cursor focuses on a hierarchical multi-agent architecture (Planner-Worker-Judge) to handle massive, parallelized coding tasks, Anthropic prioritizes "memory continuity" for single agents by externalizing state through Git history and progress logs. The piece outlines divergent yet effective philosophies for overcoming context window limitations and summarizes common failure modes in complex task execution. It is an essential read for anyone tracking the evolution of AI-driven software development.

This article captures a profound dialogue between Unmesh, Martin Fowler, and Rebecca Parsons on the essence of programming. The central thesis posits that programming is not merely translating requirements into syntax, but building systems resilient to change. It explores managing cognitive load, the intertwining of "What" (intent) and "How" (implementation), and using TDD to drive design. Furthermore, it critically assesses the role of LLMs, framing them as "translation layers" rather than architects. For developers seeking to maintain their edge in the AI era, understanding how to manage complexity through stable abstractions remains the most critical skill.

This article explores the "Desktop Agent" trend sweeping the tech world in early 2026, focusing on a deep dive into MiniMax Agent Desktop. Through three practical scenarios—organizing 400+ ebooks, encapsulating literary translation SOPs, and building a Xiaohongshu content pipeline—the author demonstrates how AI leaps from "advisor" to "executor" by accessing local context. A key insight is how Expert Agents transform personal expertise into reusable digital assets.

This article provides a comprehensive overview of the Coze 2.0 upgrade, ByteDance's leading AI platform. Coze 2.0 marks a strategic shift from a "reactive tool" to a "proactive partner." Key highlights include Agent Skills, which encapsulate industry best practices into reusable capabilities; Agent Plan, enabling agents to autonomously execute long-term goals and report progress; and Coze Coding, a Vibe Coding platform that simplifies app development.

Li Auto's SVP Fan Haoyu delves into the logic behind their first AI hardware, Livis. Diverging from mainstream AR paths, Livis adopts a restrained approach: 36g weight, no-display design, and a self-developed RTOS on a low-power MCU architecture, prioritizing high availability as an "always-on companion." Fan also shares his "Hands in the Mud" philosophy and the "6211" time management framework, discussing how Li Auto expands from its automotive "Island" to an AI-driven "Continent."

As AI evolves from generative to agentic, UX design is undergoing a paradigm shift from simple usability to trust and accountability. This article explores the core characteristics of AI agents—autonomous reasoning and planning—and proposes a taxonomy of four autonomy levels: Observe-and-Suggest, Plan-and-Propose, Act-with-Confirmation, and Act-Autonomously. Author Victor Yocco provides a practical research playbook for developers and PMs, covering methods like mental-model interviews, agent journey mapping, and simulated misbehavior testing. By establishing metrics like intervention and rollback rates, the piece offers a clear roadmap for building controllable, transparent, and ethical AI agent systems. It is an essential resource for technical leaders exploring the boundaries of AI automation.

This podcast features an in-depth interview with Zhang Yueguang, the founder of the viral "Miaoya Camera." He reflects on his entrepreneurial journey after leaving ByteDance and Alibaba, offering a provocative thesis: Miaoya is not a true AI Native product, but rather an AI-enhanced Internet product. He argues that the AI era demands a paradigm shift from "Process-oriented design" to "Context-oriented design," where PMs focus on optimizing interaction boundaries rather than deterministic paths. The discussion covers his "One Way Door" product philosophy and his latest ventures in AI Otome games and Agent tools (Dokie).

This insightful piece synthesizes the a16z AI Apps team's strategic perspective on the AI cycle, framing it as the most rapid platform shift since PC, Cloud, and Mobile. It highlights three pivotal investment themes: the AI-native transformation of existing software categories, "Software as Labor" (where software replaces human tasks at higher price points), and "Walled Gardens" built on proprietary data moats. Through case studies like Eve and Salient, the authors argue that in an era of "Vibe Coding," defensibility is no longer about the code itself but about owning end-to-end workflows and proprietary data loops.

Can a non-technical PM build real apps? Meta's Zevi Arnovitz proves it's possible through a systematic AI workflow. Using Cursor and Claude, he automates development with "slash commands" and solves the code review hurdle via multi-model cross-validation. This session is a high-density guide for anyone looking to bridge the gap between product vision and technical execution using modern AI agents.

In this insightful All-In podcast episode, Microsoft CEO Satya Nadella explores how AI is reshaping Microsoft and the global business landscape. He introduces the "Macro-delegate and micro-steer" paradigm, explaining how Microsoft achieved massive revenue growth while maintaining a flat headcount through structural shifts like "Full-stack builders." The discussion covers the evolution of AI Agents, the strategic logic behind the OpenAI partnership, and the global diffusion of the US tech stack.