What it does

The authoring canvas is where you create and shape an agent: give it a name and description, write instructions, attach knowledge, add topics and tools, then test it in the side panel before publishing.

Key facts

• You can spin up an agent by describing it in natural language, then refine, or start from a blank canvas.
• The agent's name, description and instructions all feed the model — vague instructions produce vague behaviour.
• The Test pane runs live against your draft as you build; changes take effect without a publish.
• Nothing reaches users until you publish, and publishing is per-channel.

When to use / skip

This is the starting point for every custom agent. If you only need generative answers over a document set, you can stay almost entirely on this canvas and never touch a classic topic.

Configuration decisions

• How much you rely on instructions and orchestration vs hand-built topics.
• Which knowledge sources and tools to attach up front.
• Naming and description conventions so agents are findable and governable later.

Gotchas

• The description isn't just a label — orchestration uses it to decide when the agent or its actions fire. Write it deliberately.
• Testing against the draft is not the same as testing the published channel. Verify in the real channel before sign-off.

Consultant notes

• Set instruction-writing standards early; it's the single biggest lever on quality.
• Keep a template agent with your standard guardrails and disclaimers so makers don't start from scratch.
• Put agents in solutions from the first build — retrofitting ALM later is painful.

---

Recheck when the authoring experience updates.