The Agent Loop

Module 6, Lecture 6.1 | Building the Coding Agent

This lecture takes a deep dive into the two-loop architecture that powers every tool-calling agent. Building on the coding agent introduced in Lecture 5.3, it examines why two loops are necessary, how stop_reason drives the inner loop as the sole branching condition, and how the messages array grows through single-tool and multi-tool exchanges. Students live-code the agent loop from scratch using stub tools, tracing the messages array at each step to see exactly what the model sees — and why tool results enter the conversation as role: "user".

Read the full lecture narrative

Additional Resources

• Lecture slides
• ReAct: Synergizing Reasoning and Acting in Language Models — The foundational paper (Google Research, ICLR 2023) introducing the ReAct pattern of interleaving reasoning and tool actions. The conceptual basis for the inner loop's reasoning-acting cycle.
• Building Effective Agents — Anthropic's canonical guidance on agent architecture. Covers the simple while-loop pattern, stop conditions, and tool design principles demonstrated in this lecture.
• Effective Context Engineering for AI Agents — Anthropic Engineering's treatment of how the messages array grows over a session and how tool results enter as role: "user". Bridges directly to the context growth problem introduced at the end of this lecture.
• The Canonical Agent Architecture: A While Loop with Tools — A practitioner-focused explanation of the minimal while-loop pattern, stop_reason branching, and tool_use_id from an industry engineering perspective.
• Anthropic Tool Use Documentation — The official API reference for tool schemas, tool_use and tool_result content blocks, and stop_reason values. The specification that the inner loop implements.