Professional architect sharing high‑quality architecture insights. Topics include high‑availability, high‑performance, high‑stability architectures, big data, machine learning, Java, system and distributed architecture, AI, and practical large‑scale architecture case studies. Open to ideas‑driven architects who enjoy sharing and learning.
This article dissects Claude Code's memory loading chain by examining its source code, revealing how CLAUDE.md, Auto Memory (MEMORY.md), and live workspace evidence interact — and why memory files are navigation aids, not enforceable configuration.
This article analyzes Pi's minimalist agent harness architecture, contrasting its four-tool core with Claude Code, Codex, and DSH, detailing its session tree, dual-loop execution, context compaction, and the trade-offs of pushing workflows to extensions for greater runtime controllability.
This article traces the Loop architecture lineage from Universal Transformer to recurrent depth, examines engineering challenges like placement, state drift, dynamic stopping, and KV cache, and argues that as internal computation decouples from parameters and visible reasoning, Agent verification must rely on external evidence like tests, approvals, and versioned artifacts rather than chain-of-thought alone.
The article debunks claims that vector databases are obsolete, analyzing how Claude Code and Cursor integrate retrieval into agent runtime loops rather than abandoning RAG, and proposes a five-layer architecture where vector indexes serve as retrieval projections alongside grep, semantic search, and authoritative sources.
The article analyzes Codex CLI's experimental context management system (v0.153.0), which replaces monolithic compaction with four distinct state types—current working set, handoff notes, searchable history, and external facts—detailing the model-driven window-switching protocol, token budget exposure, harness fallback mechanisms, and recovery considerations for long-running coding agents.
The article explains why swapping the Harness — the runtime environment around an AI model — changes agent behavior even when the model weights stay identical, covering system prompts, tool definitions, agentic loops, translation layers, execution boundaries, feedback fidelity, context compression vs. immutable event logs, and architectural trade-offs illustrated by Pi, Codex, DSH, and the claudex experiment.
This article analyzes how DeepSeek Harness achieves 99% prompt cache hit rates by enforcing deterministic request assembly, showing that cache hit rates serve as a diagnostic tool for runtime stability rather than just a cost metric.
DeepSeek Harness (DSH) re‑architects the agent runtime by separating capability loading, clean exit handling, and task recovery into distinct boundaries, introducing Profile‑Preset configuration, a live Cordis runtime graph, a unified tool pipeline, session event logs, and dynamic Cordis for on‑the‑fly capability changes.
The article dissects where an AI agent loop can pause—model stream end, tool batch return, turn completion, driver activity, and long‑term goal—by examining DeepSeek Harness and Pi implementations, their code structures, stop‑reason handling, logging, and practical trade‑offs in production scenarios.
The article analyzes Claude Code’s /loop and /goal commands, explaining how /loop schedules recurring checks while /goal defines completion criteria, detailing their syntax, durability options, state management, failure handling, and practical use cases for automating SDK upgrades, PR creation, and CI monitoring, and discusses how agent architecture should evolve as models become more capable.
