Build 3 Production AI Agents in Python with AgentSpan: Memory, RAG, and Orchestration

Key Takeaways

• Production agents need seven things: durability (recover without restarting), retries, human-in-the-loop, observability, long-running task support, scale and testing.
• AgentSpan splits the system into a worker (your code) and a server (provided, open-source) that tracks state, history and orchestration — so a crashed run reconnects and resumes instead of losing work.
• API keys live on the server, not in worker code, which is more secure; you export the provider key (OpenAI, Anthropic, Gemini, etc.) before starting the server.
• Tools are plain Python functions wrapped with an @tool decorator; the function name becomes the tool name and the docstring becomes its description the model reads.
• Adding ConversationMemory(max_messages=50) gives the agent recall across turns; the server dashboard shows every tool call, input, output and stop reason in real time.

Commands & Code Mentioned

pip install agent-span      # or: uv add agent-span
uv run agentspan doctor     # check the install (Java 21, disk, API key, server jar)
uv run agentspan server start
export OPENAI_API_KEY=<your key>   # set before starting the server
uv add python-dotenv pydantic firecrawl-py
@tool   # decorator that turns a Python function into an agent tool
uv run agents/agent1.py