A program manager oversees a mid-size SaaS portfolio with strict audit needs and a product squad that wants a tailored assistant trained on their PRD template, Jira fields, and tone for stakeholder updates. They ask whether to assign work to a Copilot cloud agent or build a custom agent. Which distinction should guide their choice between Copilot cloud agents and custom agents?

The correct distinction reflects platform division of labor: Copilot cloud agents run as autonomous GitHub-hosted workers that open branches, execute code, run tests and raise pull requests with permissions and audit trails, while custom agents encapsulate reusable tailored instructions, knowledge and tool bindings for domain-specific communication such as the PRD template, Jira field mapping and stakeholder tone. The generic-Q&A description fails the customization concept because it retains cloud coding but reduces the custom build to untailored answers, removing the templates and tool bindings that justify building it. The local-IDE-chat description fails the cloud-execution concept because while it captures PRD and Jira tailoring, it confuses interactive editor help with independent hosted compute. The reversed-roles description fails separation of concerns by assigning tailored tone and Jira tools to cloud and autonomous hosted coding to custom, inverting each type's strength. Exam caveat: learning tone or accessing Jira alone does not determine type since both can retrieve context; the decider is reusable tailored behavior versus autonomous GitHub execution with logs. Practical check: confirm where work must run and be audited and whether output must follow squad templates, then map hosted autonomous execution to cloud and reusable tailored instructions plus tools to custom, validating with a pilot pull request and tone-checked update.