Column

Agent System Design Anatomy

Control, state, identity, and responsibility in real systems

3 essays · 58 min total

Systems called “agents” give models very different kinds of control. A coding kernel, a workflow engine, and a persistent messaging gateway place their state, permissions, and recovery responsibilities in different layers.

This series studies those boundaries through pinned source versions. Each case separates observed implementation from interpretation and unknowns, with an editable architecture diagram and a concrete failure scenario.

The following case studies are available in English:

  1. Pi: the minimal kernel and the responsibilities it leaves to the host .
  2. n8n: workflow control, queues, and external side effects .
  3. OpenClaw: persistent gateways, session identity, and trust boundaries .

The complete Chinese series contains eleven essays. These English editions keep their original part numbers so readers can move between the two languages without losing their place in the series.

Contents

Best read in order
  1. 18 min read

    Pi by Subtraction: What the Minimal Agent Kernel Keeps and Who Takes Responsibility

    Trace Pi v0.84.1's agent loop, default tools and JSONL sessions to see what a minimal kernel keeps and who takes over security, recovery and workflow decisions.

  2. 21 min read

    n8n's Deterministic Exoskeleton: Why Queues and Workers Do Not Guarantee Exactly-Once Side Effects

    Trace n8n 2.33.6 across Redis queues, workers and Postgres to see why scaling workflow throughput does not guarantee exactly-once effects in external systems.

  3. 19 min read

    OpenClaw's Long-Running Gateway: Continuity Does Not Mean One Shared Session

    Trace OpenClaw v2026.7.1-2 through admission, bindings, agents and session keys to see how identity dimensions shape continuity, isolation and permissions.