Custom Agents
Custom Agents let you build and manage your own Smart Engineering agents. An agent has a system prompt, a model configuration, a set of tools and capabilities, optional pinned skills, recipes, and reference packs, and limits such as the maximum number of reasoning rounds. Published agents become available in Agentic Chat.
Open the agent catalog from Agentic → Chat → Personalize → Agents.
When to use
- Create a specialized assistant for a recurring job, such as reviewing a specific kind of legacy code.
- Give an agent a fixed model, tool set, and reference context so results are consistent.
- Package expertise (skills and recipes) behind a single agent your team can chat with.
Agent lifecycle
| Status | Meaning |
|---|---|
| Draft | Editable. Not offered to new conversations. |
| Active | Published and selectable in Agentic Chat. |
| Deprecated | Hidden from new conversations. Can be restored to draft. |
From the catalog you can create, edit, duplicate, deprecate, restore, chat with, run in the background, or permanently delete an agent (deleting also removes its executions).
Create an agent
- Open Personalize → Agents and select New agent, or Create with AI to generate a draft from a description.
- Set the name (a slug is derived from it), description, and version.
- Write the system prompt and the user prompt template.
- Choose the LLM provider and model, temperature, and maximum output tokens.
- Review tools and capability bundles. New agents start with all tools and capabilities enabled; narrow them with Customize tools when the agent should be restricted.
- Optionally pin skills, recipes, and reference packs the agent should always have available.
- Set execution limits: maximum rounds, maximum tool calls per round, and a timeout.
- Define output artifacts the agent is expected to produce, if any.
- Set visibility (private or organization) and save. The agent is created as a draft.
Run an agent
- Chat: open a conversation with the agent in Agentic Chat.
- Run in background: start an execution with a single input and review it later.
Executions
Personalize → Agent executions lists background runs with their status (created, running, completed, failed, cancelled). Open an execution to review the rounds, tool calls, token usage, final response, and any generated artifacts. You can cancel a running execution or delete a finished one.
Expected result
A published agent that appears in Agentic Chat with the prompt, model, tools, and reference context you configured, plus a record of every background execution.