Anthropic's Claude is both the tool Pebblous uses every day and the model we analyze most often. As a tool, it runs inside our own pipelines like NanoClaw and dc-story-produce to generate content. As a subject of analysis, we track how the model is built, what it is allowed to do, and where it breaks. Claude Watch is the hub where both vantage points meet.
The five chapters of this series operate at different layers. The politics of model release (why Anthropic won't release Mythos), the harness postmortem (the real reason Claude got dumber), AI alignment (the Bernie Sanders interview and the sycophancy problem), agent architecture (the Claude Agent SDK 5,526-line dissection), and coding behavior correction (Karpathy's coding pitfalls and CLAUDE.md). Above the model, inside the model, below the model, beside the model — Claude seen from multiple angles.
The question Pebblous keeps asking from this vantage point is one. The quality of AI judgment is, in the end, a function of data. When sycophancy aligns a model with user beliefs, when a harness determines behavior through system prompts, when an Agent SDK chooses how to call tools — behind each lies the question of what data made the model and what data enters its inference. The vantage points of DataClinic and Neuro-Symbolic × Ontology converge here.
The model-release chapter. Why Anthropic keeps Claude Mythos — a model with zero-day cybersecurity capability — behind closed doors. What Project Glasswing means and how responsible disclosure is redefining model governance.
The harness-postmortem chapter. The model didn't change, but the answers got worse — what's actually happening. A dissection of how system prompts, context management, and tool-call layers shape model behavior, based on Anthropic's official analysis.
The AI-alignment chapter. The moment Claude's position shifted live on air, used as an entry point into how RLHF-aligned models echo user beliefs (sycophancy) and into the political stakes of AI regulation.
The agent-architecture chapter. How Pebblous designed NanoClaw on top of Claude Agent SDK and Managed Agents, and where container isolation, tool-call security, and context management actually break — read straight from 5,526 lines of open source.
The coding-behavior-correction chapter. Starting from the six LLM coding pitfalls Andrej Karpathy identified, how CLAUDE.md can correct agent behavior — tracing the causality backwards: when data quality collapses, code collapses with it.