Inside Claude Code: How Anthropic’s Agentic Harness Powers Next‑Gen AI Agents

01 It’s Not a Prompt, It’s an Operating Model

Claude Code’s system prompt is not a static block; the SYSTEM_PROMPT_DYNAMIC_BOUNDARY separates cacheable static prefixes from session‑specific dynamic suffixes, allowing Anthropic to treat prompts as orchestrated runtime resources and dramatically reduce token costs.

02 Institutionalized Good Behavior

The code embeds AI‑engineer best‑practice rules directly into the model via the SimpleDoingTasksSection , enforcing safe actions, confirming destructive operations, and applying a “blast radius” mindset that minimizes variance and prevents uncontrolled modifications.

03 Context Is a Scarce Resource

Numerous optimizations protect context, including cache boundaries, shared prompt caches for forked paths, on‑demand skill injection, dynamic MCP instructions, and transcript/resume mechanisms, all designed to maximize cache hits and minimize token waste.

04 Agent Specialization, Not a Universal Worker

Claude Code defines distinct built‑in agents: Explore Agent (read‑only code exploration), Plan Agent (step‑by‑step implementation planning), and Verification Agent (aggressively attempts to break the implementation), illustrating a specialization‑first architecture.

05 Scheduling Chain: AgentTool → runAgent → query

The scheduling chain acts as a full‑featured orchestration controller, handling permission filtering, MCP dependencies, worktree isolation, and telemetry, while runAgent constructs a complete sub‑agent runtime with hooks, permissions, and tool sets.

06 Skills, Plugins, Hooks, MCP: Model‑Aware Extensions

Skills are prompt‑native workflow packages, plugins combine prompts, metadata, and runtime constraints, hooks provide a governance layer that can modify inputs or block execution, and MCP delivers both tools and usage instructions, turning it into a behavior specification channel.

07 Tool Execution Pipeline

Tool execution follows a runtime pipeline rather than direct calls; pre‑tool hooks can rewrite inputs, enforce permissions, or halt continuation, ensuring that security policies are never bypassed.

Conclusion: The End Goal of the Agentic Harness

The Harness unifies prompt architecture, tool runtime, permission models, agent orchestration, skill packaging, plugin systems, hook governance, MCP integration, and context hygiene into a single, product‑grade system, demonstrating that the future of AI agents lies in comprehensive system design rather than ever larger models.

https://github.com/instructkr/claw-code
https://github.com/tvytlx/claude-code-deep-dive