Agentic loops

• Each pass has four stages: send the request, inspect stop_reason, run the requested tools, return the results for the next pass.
• tool_use means keep going; end_turn means the task is finished.
• Tool results are appended to the message history so the next call can reason about them.
• Claude chooses the next tool from context; a fixed tool sequence or decision tree only fits steps that never vary.
• A reply can contain text and a tool request at the same time, so text alone never means done.
• An iteration cap is a safety net, not the normal way to stop.

• Write control flow that continues on tool_use and terminates on end_turn.
• Add tool results to the conversation between passes so the model can use new information.
• Recognise the three anti-patterns: parsing natural language to end the loop, using an arbitrary cap as the main stop, and treating assistant text as a completion signal.

• "If the reply contains a sentence, Claude is finished." Only the stop reason says that.
• "Sending just the latest message is enough." Without the history Claude repeats lookups and loses earlier findings.
• "A fixed pass limit is a fine stopping rule." It cuts real tasks short and wastes passes on short ones.

Loops that end on text content, stop on a pass count, or script the tool order instead of letting the model decide.

• 1.4Workflow enforcement and handoff
• 2.2Structured error responses
• 5.1Conversation context management