How-to
Create an Agent
The Agent builder is a 9-step wizard. Each step persists as you go, so you can stop and resume. By the end you have a published, versioned Agent ready to test and ship.
What you'll learn
- What each of the 9 builder steps configures
- Which steps are required and which are optional
- How LLM, Tool, and Guardrail choices interact at runtime
- How to publish and version an Agent
Before you start
Open Agents in the sidebar and click + Create Agent. The wizard opens on step 1. You can navigate back and forth — nothing is committed until you click Publish on step 9.
Walk through the 9 steps
- 1
Step 1 — Basic Info
Give the Agent a clear name (e.g. "Support Triage Agent"), a one-sentence description, and a category. Categories: DevOps, Security, Quality, Support, Productivity, Analytics, or Custom. - 2
Step 2 — LLM Selection
Pick the model that backs the Agent. Choices include OpenAI, Anthropic, Google, AWS Bedrock, Azure OpenAI, and self-hosted models. Configure providers under Settings if the list is empty. - 3
Step 3 — Tool Selection
Tick the integrations this Agent can invoke at runtime — Slack, Jira, Zendesk, Postgres, GitHub, and so on. Only enabled Tools appear; connect new ones from the Integrations cluster first. - 4
Step 4 — Skill Selection
Attach one or more Skills — reusable, versioned prompt packs from the Skill library. Skills let multiple Agents share specialized behavior without duplicating prompts. - 5
Step 5 — RAG Configuration
Bind Knowledge Bases the Agent can retrieve from. Pick one or many; choose top-k and a similarity threshold. Skip this step for Agents that do not need document grounding. - 6
Step 6 — Memory Configuration
Pick a memory mode. Short-term keeps state inside a Session; long-term persists across Sessions; hybrid does both. Default is short-term. - 7
Step 7 — Voice Configuration
Enable voice if you want speech in or speech out. Choose an STT and a TTS provider. Leave disabled for text-only Agents. - 8
Step 8 — Guardrails Configuration
Attach a Guardrail profile. The profile inspects every prompt, Tool call and response. Start with the Standard preset; tighten or relax from there. - 9
Step 9 — Review and Publish
The review screen lists every choice. Click Publish to create a versioned Agent. You can now Test it, expose it as an API, or drop it into a Workflow.
Frequently asked questions
- What happens if I close the wizard halfway?
- Your progress is saved as a draft. Reopen the Agent and you resume from the step you left.
- Can I change the LLM after publishing?
- Yes. Edit the Agent, change the LLM, and publish again. The change creates a new version — old Runs remain attributed to the old version.
- Do I need a Guardrail profile to publish?
- No, but it is strongly recommended for any Agent that goes beyond internal testing. The Standard preset covers the common cases out of the box.
- How many Tools can one Agent use?
- There is no hard limit, but Agents perform best with a focused Tool belt — typically 3 to 8. Past that, consider splitting the job across multiple Agents in a Workflow.