First Agent with Retrieval and Tools

This tutorial is the anchor journey for the docs program. It is intentionally opinionated and narrow.

Outcome

By the end, you should have:

one runnable agent graph
one retrieval source that can materially change an answer
one tool with a visible runtime effect
one inspectable run you can reason about afterward

Set the runtime floor

Choose one tenant, one model, one small knowledge scope, and one tool. Keep all four fixed through the first run.

Create the agent graph

Build the smallest graph that can still prove the platform path:

Start
agent node
retrieval/tool integration path as needed
End

Connect retrieval

Use knowledge that answers a question the base model should not answer reliably without grounded context.

Connect one tool

Use a tool whose execution is easy to confirm. If you cannot tell whether it ran, it is the wrong first tool for this tutorial.

Run with one verification prompt

Use a prompt that should fail or degrade when retrieval or the tool is missing.

Inspect the run

Check whether:

retrieval context appeared
the tool was callable and actually called
the final answer reflects those intermediate steps

Acceptance check

the agent run completes
retrieval affects the answer
the tool path is visible
the output can be reviewed after the run

Verification checklist

tenant access confirmed
model configured
knowledge path available
tool attached and visible
run trace inspectable

Known gaps

This tutorial is ready as the base public path, but the team still needs to execute a fresh full run against the live product and convert any friction into tracked gap-log items.