Create and manage Sub Agents

Prerequisites

agents.read to inspect/export agents; agents.write to create, edit, enable/disable, delete, import, or use AI design. The current panel disables even export controls without write permission.
chat.use to launch a chat. Agent definitions are deployment-wide, not isolated per application tenant.
A tested AI provider for draft and enhancement generation.
Tested, enabled connectors for connector-backed tools.
settings.read for the provider/model pickers and saved tool-policy diagnostic; Azure/Entra permissions remain separate from these application capabilities.

Route

Open /automations/agents.

**Built-in Sub Agent library and review-mode instructions.** The native security-specialist definitions are reusable instructions, not generated reports. No agent was enabled, disabled, enhanced, exported, or invoked during capture, and no provider, OAuth, or Azure connectivity was verified. Synthetic example data; names, identifiers and measurements do not describe a live tenant. View full-size screenshot.

How to create and verify a Sub Agent

Select New agent and enter a clear name, category, and instructions. Alternatively, select Generate with AI, describe the goal, and submit answers with Continue; Generate now skips questions and omits unsubmitted current-step answers.
Select the intended provider/model or defaults. The editor does not expose a connection selector; review imported connection references and the actual chat scope separately.
Choose Review for initial validation. Autonomous removes the interactive gate for write-classified tool calls; it is not just a faster review mode.
Search/filter Connector tools and select the minimum needed. These are tool-name selections, not a unique choice of destination when several connectors offer the same tool.
Turn off complete Azure catalog routing and select appropriate bundles/exact names when a narrow scope suffices. Opt into Entra only when needed through its bundles/names or complete-catalog control. Review retained policies, not only the visible checkbox.
Select Save. New agents default to enabled; use the card switch to disable an agent awaiting review. Enable after review and select its Chat action, or choose it from the chat picker.
Issue a harmless read-only verification prompt and inspect actual connection, tool activity, and response.

Expected result: The enabled agent appears in the chat picker and responds using its saved configuration.

Verification: Confirm the selected agent name, target scope, and actual tool calls. A saved tool-policy preview is diagnostic, not a live test of unsaved changes or proof that every eligible tool was initially exposed.

How to enhance or disable an agent

Pause schedules using this agent before changes that could alter their behavior. Select Edit for direct changes or Enhance for the assessment/interview.
Submit answers with Continue, or use Enhance now to skip remaining questions. Inspect What changed, run-mode changes, and Compare before / after.
Select Review & save to open the normal editor; this does not yet persist the draft. Review instructions/tools again, then select Save. Regenerate makes another draft; Cancel leaves the stored definition unchanged.
Disable the agent when it should disappear from new-chat selectors, and clear it from affected existing chats. Disable does not halt scheduled or existing-chat references.
Delete only after schedules and chat references have been reviewed. Confirm Delete knowing there is no trash/restore action; a missing scheduled agent can fall back to default behavior rather than stopping.

Expected result: Saved edits affect later turns/runs; disabled agents disappear from launch selectors without being deleted. Independently paused schedules do not automatically dispatch further work.

Verification: Reopen Edit to verify persistence, check new-chat selector behavior, and inspect referencing schedules explicitly. Disabling the agent alone is not verification that automation stopped.

How to change models for selected agents

Select the intended agent checkboxes. Select all includes the whole library, even when a category filter is active; check the selected count.
In Set model for N, choose the provider and model or provider default.
Select Apply to N, then reopen the affected agents to verify model settings and preserved instructions/tool policies.
If an error occurs, inspect each selected agent before repeating the operation: updates are separate requests and can partly succeed.

Expected result: Selected agents receive the new provider/model without replacing their saved instructions.

Verification: Check a later controlled chat turn. In chat, an agent forces its model only when both provider and model are set; otherwise chat/global model selection applies.

How to export and import agents

Use a card’s Export, Export selected, or Export all. The bundle contains instructions, provider/model, connection reference, tool policies, run mode, and enabled state, but not category or server IDs/timestamps.
Inspect the definition and remove secrets/environment identifiers before sharing. Connector credentials are not packaged, but the instructions and references are not automatically sanitized.
Pause affected schedules on the destination first. Review naming and enabled state before import: the UI immediately updates an existing agent with the same name, rather than opening a staging/preview dialog. Use a distinct reviewed name for a separate copy; set enabled false in a reviewed import when it must arrive disabled.
Select Import and choose the reviewed single/bulk bundle. Read the created/updated counts. Unknown connector tools are dropped; records without a meaningful name/instructions are skipped or rejected.
Open the resulting agent and review model, run mode, MCP bundles/names, connector tools, and category. Use Export to inspect its stored connection reference, which has no editor field. Reassign category manually because it is omitted from export.
After review, enable the agent and perform a harmless read-only chat check before re-enabling its schedules.

Expected result: Reviewed definitions are created or updated by name without importing connector credentials. Enabled state follows the imported value, defaulting to true.

Verification: Open the imported agent, verify model and tool mappings, and run a harmless read-only prompt.

Safety and rollback

Creation, edits, imports, and enablement change application configuration. AI drafting/enhancement calls the configured provider; execution can invoke Azure/external tools. Review mode gates classified writes, but Autonomous explicitly removes that interactive gate. Downstream authorization is still required.

There is no agent-definition version history, restore, execution-cancel, or retry-failed control. Reimporting a previous reviewed definition may update a same-named agent but does not undo prior external effects or restore deleted IDs. Pause schedules and clear affected chat selection before disabling/deleting; never rely on the enabled flag as a runtime kill switch. Inspect each result after bulk operations or imports because partial changes can persist.

Troubleshooting

Symptom	Cause and resolution
Generate action fails	A provider failure or unusable draft response can stop generation. Check the active provider and simplify the goal; wizard Retry regenerates a draft, not a prior execution.
Imported tools are unavailable	Configure and enable the matching connector, then edit the tool selection.
Agent behaves too broadly	Pause its schedules and stop using it in existing chats, then reduce instructions and tool policy. Disabling only hides launch choices. Verify with a read-only prompt after review.
User cannot edit	Assign `agents.write`; `agents.read` is intentionally view-only.
Model/diagnostic choices absent	Those requests use `settings.read`; ask an administrator to check access and active provider availability.
Import replaced an agent or lost category/tools	Same-name overwrite is the UI default; category is not portable and unknown connector tools are dropped. Inspect all imported settings and restore only reviewed values.
Draft ignores the current answers	Generate now/Enhance now use previously submitted answers. Submit current answers with Continue before skipping remaining questions.