Quickstart

This is the shortest useful path through the platform. The goal is not full mastery. The goal is to prove one agent can read context, call a tool, and produce a traceable run.

Create or access a tenant

Use a tenant where you can configure models, knowledge assets, and tools. If tenant bootstrap still depends on internal setup knowledge in your environment, treat that as a docs gap immediately.

Configure a model

Set one default model you trust for normal text generation before building the agent. Do not start with multi-model routing in the first pass.

Prepare one retrieval source

Create one knowledge input that can answer a narrow set of questions. Keep the corpus small enough that you can tell whether retrieval actually helped.

Attach one tool

Choose one tool with an obvious success condition. A broad tool catalog hides product gaps. A single tool makes run inspection clearer.

Build the first graph

Start with a minimal flow:

Start
one agent node
retrieval connection if needed
one tool path if needed
End

Run and inspect

Execute the graph with one test prompt you can verify manually. Then inspect the run trace, tool usage, and outputs before expanding scope.

Recommended first prompt

Use a prompt that forces both retrieval and tool usage to matter. If the answer is easy without either, you will not learn much from the run.

Success criteria

the run completes without hidden manual intervention
you can tell which context came from retrieval
you can tell whether a tool was called
the output is inspectable after the run

Verification checklist

tenant is accessible
model is configured
one retrieval source is available
one tool is visible to the agent
run can be executed and inspected

Known gaps

This quickstart is seeded from canonical platform docs, but the public end-to-end setup path still needs a full fresh-run validation in the live product flow.