Agent Loop

The agent loop is the fundamental operating pattern of tool-using AI agents. A prompt plus settings enter the loop; the model decides whether to answer directly or call a tool; an approved tool executes and returns a result; the result re-enters the loop. This repeats until the agent produces a final output.

How it works

Prompt + Settings → Model → Needs Tool?
   ↑                           |
   |           No → Final Result
   |           Yes → Permission Check → Tool Execution → Tool Result → (back to Model)

Every tool call is mediated by a permission check that determines whether the tool can run automatically, requires human approval, or is blocked. This is the governance boundary that transforms a capable model into a production-grade agent.

Why the loop boundary matters

• Tools convert reasoning into action — without them, the model can only produce text.
• Permissions convert action into governed operation — without them, any tool can run unchecked.
• Sessions convert one-off work into resumable continuity — session IDs and result metadata allow recovery from interruption.

The key design insight: the loop is not just the model. It includes the settings that constrain it, the hooks that intercept it, and the session that preserves its state.

Enterprise design rules

1. Start with read-only tools; add write access only after logging and approval hooks are in place.
2. Use hooks at every tool invocation for audit and policy.
3. Track session_id, token usage, cost, and result status for every loop run.
4. Validate outputs — do not assume a completed loop means a correct result.

Related

• BoundedAgent — pattern for constraining which tools the loop can reach
• AgenticGovernance — hooks, logs, and approval flows that govern the loop
• ContextBudget — the finite working memory consumed during each loop iteration
• ClaudeAgentSDK — the SDK that embeds this loop in custom applications
• MCP — protocol for exposing external tools to the loop