Claude Code reveals a hard truth about AI agents. The model is only a small part of the working system. A public analysis of the codebase found that about 1.6% of Claude Code handles AI decision logic. The remaining 98.4% runs the infrastructure around the model: permissions, memory, tools, recovery, and activity records.

This infrastructure is the harness: the control layer around the model. It decides what the AI accesses, what the AI changes, which tools the AI uses, and what happens when work breaks. The model reasons about the next step, and the harness turns reasoning into safe action.

The core loop stays simple. The paper describes a loop that calls the model, runs tools, and repeats. The AI receives context, picks the next action, gets the result, and continues. Most of the operational system surrounds this loop.

The paper shows how much structure sits underneath. Claude Code uses seven permission modes that define approval rules for different actions. The system also runs context compaction, shrinking long work histories without losing the useful parts. These systems help the AI continue working without losing track of the task.

This is a business issue. A model writes a strong answer. An agent completes work across real systems. That work touches files, code, customer data, internal documents, or outside services. Every weak control creates cost. A wrong action creates rework, risk, or delay.

The common mistake treats autonomy as the goal. More freedom does not automatically create more value. A hands-off agent with weak controls moves fast in the wrong direction. The agent uses the wrong data, loses track of a task, alters the wrong file, or produces output outside your review process.

Claude Code points to a stronger operating model: the model acts like a consultant inside a controlled operation, proposing actions while the wider system checks what is allowed, preserves history, connects tools, and handles mistakes.

Cost control begins here. Permission rules reduce damaging actions. Memory systems reduce repeated work and lost instructions. Recovery systems reduce the cost of failure. Logs make work easier to review later. These controls determine whether your AI system runs at scale without building hidden operational debt.

The same logic applies to model choice. If your business owns the harness, you swap models without rebuilding the working system. You keep control over cost, performance, and vendor dependency. Models improve underneath the harness while your rules and records stay stable.

Data governance works the same way. The harness decides what information reaches the model, what stays internal, what gets logged, and what needs approval. As agents touch more systems, this becomes a core operating requirement.

Your action plan is direct. Define what the agent can access, what it can change, when approval is required, how context is handled, which tools are available, and how failures are recovered before expanding what the agent does on its own.

Sources

arXiv, Dive into Claude Code: The Design Space of Today’s and Future AI Agent Systems — arxiv.org
VILA Lab, Dive into Claude Code GitHub Repository — github.com