Building the Coding Agent System Prompt

Module 5, Lecture 5.3 | Section 3: Prompt and Context Engineering

This lecture puts everything together: system prompt, tool schemas, and working agent code. Starting from the three starter tools (list_files, read_file, edit_file), it builds the complete system prompt section by section — identity, tool descriptions, behavioral guidelines, constraints, and a few-shot example. It then introduces two Python programs: code-agent-v0.py, a diagnostic that reveals the raw API response structure before adding loop complexity, and code-agent-v1.py, the first fully working agent. The central lesson is the nested loop architecture — an outer loop that manages the conversation and an inner loop driven by stop_reason that handles tool execution — and how the messages array grows through each iteration to form the model's complete working memory.

Read the full lecture narrative

Additional Resources

• Lecture slides
• code-agent-v0.py — Diagnostic script: single API call showing raw stop_reason and tool_use response structure, with stub tools.
• code-agent-v1.py — Working coding agent: outer/inner loop, real filesystem tools, full conversation history.
• Build a tool-using agent (Anthropic Docs) — Step-by-step tutorial building the full two-loop agent in annotated Python, directly mirroring the pattern introduced in this lecture.
• How tool use works (Anthropic Docs) — Conceptual reference covering stop_reason, tool_use content blocks, and agentic loop mechanics.
• Handle tool calls (Anthropic Docs) — Reference for tool_result message formatting, the role: "user" convention, and is_error handling.
• Writing tools for agents (Anthropic Engineering) — Engineering post on tool design principles, including how system prompt and tool definitions work together.
• Agentic loop with tool calls — Temporal AI Cookbook — Third-party Python example demonstrating the same nested loop pattern generalized beyond a single course implementation.