agent-design

After a conversation with Google Labs Product Director Jaclyn Konzelmann, Peter Yang distilled 6 takeaways on how Gemini 4 is changing the way Google builds products: judgment is scarcer than code, chores-like tasks should be designed for agents first, and understanding the user's underlying problem matters more than doing exactly what they ask.

The author argues that current AI Agent products are not yet in their final form: work-oriented agents (like Codex and WorkBuddy) and personal agents (like Grok Bot) each take shortcuts, and in the future they may converge into one main agent that knows the user and orchestrates multiple work agents, with permissions and insufficient intelligence being the main obstacles.

Anthropic's Applied AI team发布指南，指出许多企业多智能体设计只是照搬组织架构图（康威定律），并建议先建立端到端评测、再逐个折叠子智能体，仅在折叠后性能下降时才保留子智能体。

A discussion of how to design approval screens for AI agent actions, arguing that showing factual effects (what will change, reversibility, policy triggers) is more effective than narrative summaries, and offering guidance on where to place human review gates.

作者分享在构建采购 sourcing agent 时，如何划定 AI 代理与人类决策的边界，采用约 80/20 分工：代理处理重复性工作，人类保留策略、谈判与最终决策，并向社区征询如何确定这条界线。

A practitioner describes building a Claude-based crypto news trading agent that skips almost everything, scores its refusals against a random baseline, and rewrites the prompt to force low-confidence actions for data collection.

该论文提出"错误复现"（error reproduction）机制，指出在LLM智能体外部记忆中持久化抽象规则会循环放大错误，并提出Anti-Persistence记忆设计：仅累积轨迹与结果、按需临时合成指导，从而在多个基准上取得更高成功率与更低LLM调用成本。

This paper shows that accumulating more memory can degrade LLM agent performance through error reproduction, where stored mistakes regenerate themselves via new abstractions. It proposes Anti-Persistence, a method that keeps trajectories but reconstructs task-specific guidance on demand, achieving higher success and lower LLM-call cost across benchmarks.

Lee Robinson 发布了在斯坦福 CS146S 的讲座视频，讲解 Grok Bot 的工作原理，涵盖 Always-on Agent 设计、harness 设计、上下文工程以及如何保持主 agent 轻量等内容，值得一看。

Wavestone AI Lab dissected 11 coding agents — including Claude Code and Codex — and found they all share a common 7-part harness blueprint (loop, LLM layer, tools, memory, safety, orchestration, extensions), ending with a reusable 90-line harness and 18 design recommendations.

ReLiveGym introduces a diagnostic evaluation environment for long-lived LLM agents that act sparsely over simulated weeks of replayed real-world news, market, and social-media streams, showing that action-timing mechanisms and feedback use are key design axes across eight base models.

This FAST workshop paper argues that LLM-based agents should treat evidence preparation—acquiring, transforming, or probing evidence before deciding—as a first-class design capability called epistemic action, enabled by 'epistemic scaffolding' interfaces and tools.

This paper introduces a controlled evaluation framework that separates answer coverage from repeatable task specialization in LLM-generated harnesses on MATH-500, finding that additional coverage largely comes from repeated executions of the same program rather than genuine specialization gains.

The paper proposes a framework for designing proactive LLM agents around three joint principles (3T): Task Capability, Temporal Allocation, and Trust, along with a five-dimensional design space. It also introduces Proactivity-Gym, a simulation-based testbed with multi-day scenarios and persona-conditioned simulated users, evaluating 23 model-harness configurations and a 30-participant human study showing that trust declines sharply when interventions misalign even if task outcomes are correct.

本文基于PyCon China 2026的演讲，讨论如何设计AI代理的组成、运行环境和生命周期。

The article critiques human-in-the-loop systems for leading to disengaged approval and advocates for scoped authority in AI agent design, emphasizing evidence and enforcement over routine clicks.

The author investigates design challenges for AI agents to manage tool budgets effectively, focusing on preventing wasteful retries and handling API failures through budget-aware strategies.

The blog post introduces Maka's Design Blog, which discusses practices and designs from an infrastructure architect, particularly relevant for AI agents. It highlights the idea that agents are replayable histories rather than processes, with logs serving as the runtime.

The article discusses how persistent memory in AI agents can amplify prompt injection risks by storing hostile instructions as trusted context, and explores a design to mitigate this while acknowledging limitations and the need for further testing over extended periods.

The article argues that multi-agent systems are often overused in AI applications, suggesting that a single agent with good tools, strict state, and clear stop conditions can be more efficient, easier to debug, and cost-effective for many workflows.