Arky | AI Agent Deep Dive

Arky - AI Engineering & AI Agent Deep Dive
  • 0.0 (0)
  • Technology
  • Updated Weekly

Arky is the AI agent deep dive. Every day we take one AI engineering topic and pull it apart in a focused technical interview - sandboxes, agent memory, RAG, evaluation, security, and cost. For each topic we explain the architecture, the trade-offs, and what you'd actually build. Recent deep dives: The Engineering of Sandbox Agents, RAG vs File-Based Memory, You Can't Grade an Agent Like a Model. What you'll hear: - Sandboxing, isolation and agent security - DNS tunneling, microVMs, prompt injection, credential modeling - How agents actually remember - RAG, file-based memory, context management - Agent evaluation and grading - why you can't judge an agent like a model - Lifecycle, state persistence, cost engineering and LLM token economics - Production: local agents vs cloud sandboxes, control planes, snapshots Who it's for: AI engineers, ML practitioners, and technical founders building and operating AI agents. Hosted by Daniel, with technical guest Maya. Full transcripts, show notes, and cited sources for every episode.

Episodes

  1. Aug 28

    为什么 AI 助手会忘事：Harness 持续学习

    为什么改动了 AI 助手的配置之后，老任务反而开始失败，模型本身却一点没变？一篇新论文给它起了名字：Harness 层级的遗忘。Daniel 和 Maya 用大白话讲解 Harness 持续学习——冻结模型外层的四个部件、把出主意和上线分开的优化器与评估器三道门、以及那个在学得多快和留住多少之间做权衡的旋钮，并给出多个环境里的实验结果和诚实的局限。

  2. Aug 28

    Why AI Assistants Unlearn Things: Harness Continual Learning

    Why does an AI assistant start failing on old tasks after you change its setup, even though the model never changed? A new paper names it: harness-level forgetting. Daniel and Maya explain Harness Continual Learning - the four pieces of the harness around a frozen model, the optimizer/evaluator gate that separates brainstorming from shipping, and the one dial that trades how fast an agent learns against how much it keeps.

  3. Aug 24

    The Engineering of Sandbox Agents

    Daniel and Maya separate the two failure classes of agent sandboxes: the eval-harness headlines (a testing/ops problem, not a boundary defeat) from genuine substrate breakouts like Unit 42's DNS tunneling. They walk through the substrate tradeoff (V8 isolates, microVMs, gVisor), the disposable-agent plus stateful control plane architecture, why attackers chase credentials rather than the model, and why prompt injection is a perception problem no fence can fix.

  4. Aug 24

    You Can't Grade an Agent Like a Model

    Why accuracy is the wrong score for an AI agent. A 50x cost gap with no accuracy reward, a 53-point CORE-Bench swing driven by the scaffold and the grader not the model, graders as an attack surface, and the four concrete moves that turn your evals from theater into evidence.

  • 4 Episodes

About

Arky is the AI agent deep dive. Every day we take one AI engineering topic and pull it apart in a focused technical interview - sandboxes, agent memory, RAG, evaluation, security, and cost. For each topic we explain the architecture, the trade-offs, and what you'd actually build. Recent deep dives: The Engineering of Sandbox Agents, RAG vs File-Based Memory, You Can't Grade an Agent Like a Model. What you'll hear: - Sandboxing, isolation and agent security - DNS tunneling, microVMs, prompt injection, credential modeling - How agents actually remember - RAG, file-based memory, context management - Agent evaluation and grading - why you can't judge an agent like a model - Lifecycle, state persistence, cost engineering and LLM token economics - Production: local agents vs cloud sandboxes, control planes, snapshots Who it's for: AI engineers, ML practitioners, and technical founders building and operating AI agents. Hosted by Daniel, with technical guest Maya. Full transcripts, show notes, and cited sources for every episode.

Information