Claude Code Agentic Loop
How Claude Code Works is a high-fit primary-source video because it names the ordinary machinery beneath the agent metaphor. Claude presents Claude Code as a loop: the user gives a prompt, the system gathers needed context, the model returns either text or a tool call, Claude Code takes action, checks whether the action satisfied the prompt, and either stops or loops again. The video also explains that context is finite, tool use distinguishes agentic work from plain chat, and permission modes shape when the agent can edit files or run commands.
The strongest Spiralist relevance is delegated agency made inspectable. A coding agent is not only a language interface; it is a model joined to a working directory, terminal actions, file edits, search tools, context management, and human approval gates. That belongs beside AI Coding Agents, AI Agents, Tool Use and Function Calling, Context Windows and Context Engineering, Agent Tool Permission Protocol, and Agent Audit and Incident Review. The key governance lesson is that the loop should leave points where a person can add context, interrupt, steer, approve, and verify instead of merely accepting fluent output.
External sources support the product frame while narrowing its claims. Anthropic's Claude Code overview describes Claude Code as an agentic coding tool that reads codebases, edits files, runs commands, and works across terminal, IDE, desktop, and browser surfaces. Anthropic's Claude Code security documentation says Claude Code uses read-only permissions by default, requests explicit approval for edits and commands, restricts write access to the started folder and subfolders unless additional permission is granted, and warns that no protection removes the need for ordinary security practice. The NIST AI Agent Standards Initiative gives independent policy context for why agent identity, authorization, secure operation, interoperability, and auditability matter as software agents begin acting for users.
The uncertainty is straightforward. This is an official two-minute product explainer from Claude, not an independent reliability, security, usability, or productivity study. It is strong evidence for Anthropic's May 2026 vocabulary around loops, context, tools, plan mode, auto-accepted edits, and permission skipping. It does not prove that verification is complete, that users will notice bad actions before approving them, that context compaction preserves every important dependency, or that configurable permissions are sufficient for sensitive repositories, regulated software, or teams running many agents at once.