Custom AI agents
Build an agent for your own task, then evaluate how it responds and uses its tools.
Choose a custom agent or a built-in team
Use AI Agent Builder when you need your own agent deliverable and chat experience. Use Workload when you want an existing specialist team such as SEO, Social or Sales. Building a custom agent does not register another built-in Workload team.
Describe the job and its boundaries
Specify the audience, allowed topics, tone, language and what the agent should do when it cannot answer. For example: “Answer questions about our pottery classes using the supplied course guide. If a refund request needs staff review, explain the hand-off instead of promising a refund.”
Provide knowledge and examples
Include the material the agent should use and example questions with expected answers. Make it clear which facts are authoritative and how the agent should handle missing information. Update the source material when policies or products change.
Define useful tools
The current agent runtime supports knowledge search, declared public HTTPS tools and hand-off events. Describe the input and intended result for each tool. A hand-off event needs an appropriate follow-up experience; it does not automatically connect every help desk or messaging platform.
Review the working chat
The platform runs the agent behind its generated chat widget. Test a normal request, a question outside its scope, a request for missing information and a tool or hand-off case. The hosted runtime remains a dependency even when you download the generated widget files.
Inspect evaluation results
With verification enabled, evaluation examines example answers, grounding, tone and language, declared tool calls and platform probes. Read the transcripts and findings, including error and timing information. Fix incorrect behaviour in the build conversation and evaluate the revised version.
Share and maintain the result
Use the supported preview and publishing options, then keep the instructions and knowledge up to date. Review the agent with real examples from its intended task before expanding its scope.