Insight Engine¶
Put the pieces together: an agent that answers questions over a database, its tools served via MCP.
What it combines¶
This capstone stitches together everything from the earlier tutorials:
- A provider and an
AgenticLoopas the reasoning core. - A database exposed over MCP, so the agent queries data through governed MCP tools.
- Guardrails to keep queries read-only and bound what the agent can touch.
- Optionally, observability to trace each run.
# sketch — the notebook has the full, runnable version
agent = AgenticLoop(
provider=provider,
tools=pool_with_db_over_mcp, # MCP tools from a DB server
system_prompt="Answer questions about the dataset. Query, then summarise.",
guardrails=Guardrails(shell_policy="denied", network_policy="denied"),
io=AutoIO(), cwd=Path.cwd(),
)
answer = await agent.run_collect("Which region grew fastest last quarter?")
print(answer.final_text)
Full notebook¶
Follow the Colab notebook for the complete build — schema, MCP wiring, and example questions.
Next: Tutorial 6 → · SQL Agent guide