Practical Guide on how to build an Agent from scratch with Gemini 3

Abstraction: Step-by-step Python agent implementation using Gemini 3 tool-use loop

Key points:

• Core agent loop: Observe -> Think -> Act -> Observe; implemented as a while loop intercepting FunctionCall objects, executing tools client-side, and returning FunctionResponse to the model
• Agent components: Model (brain), Tools (hands/eyes), Context/Memory (workspace), Loop (life)
• Gemini 3 uses "Thought signatures" to maintain reasoning context across API calls — these must be returned verbatim in subsequent requests
• Tool design rules: clear verbose names, precise docstrings (serve as prompt engineering), return meaningful error strings (not stack traces), tolerate fuzzy inputs
• Context management: prefer just-in-time loading over data dumping; use search_users(query) not get_all_users(); for long-running agents, summarize history or use external scratchpad memory
• Safety: always add max_iterations escape hatch; use human-in-the-loop confirmation for sensitive actions like send_email or execute_code

Connections: Gemini · Philipp Schmid · AI Agents · Tool Calling · Context Engineering

Source: https://www.philschmid.de/building-agents