Configure Lifecycle Agent
Configure models, instructions, tools, approvals, and repository overrides for Lifecycle Agent from Settings.
Lifecycle administrators configure Agent behavior in Settings. Start with a small global policy. Preview the combined instructions. Add repository overrides only where they are necessary.
Agent administration is available only to Lifecycle administrators. If these settings are missing, ask the deployment owner if Agent Sessions is available.
Before you start
These items are necessary:
- Lifecycle administrator access
- At least one model provider configured by the deployment owner
- A test repository and nonproduction Environment
- Agreement on which actions need approval
Keep provider credentials in the approved deployment secret store. Do not put credentials in Agent instructions, chats, or repositories.
Configure the global policy
Open Settings. Examine each Agent section:
| Section | What it controls |
|---|---|
| Availability & models | If Agent is available and which models users can select |
| Instructions & run limits | Administrator guidance and the maximum number of agent iterations |
| Permissions & approvals | Which classes of action run, require approval, or are denied |
| Tools & integrations | Built-in tool availability and shared MCP servers |
| Agent workspaces | Workspace behavior and lifecycle |
| Runtime backends | Where new Agent workspaces run |
| Safety lists | Excluded tools and paths, plus allowed write paths |
| Session review | Audit records for Agent sessions |
Save one section at a time. Use a test session to make sure that each result is correct. A new session clearly shows the changed instructions or policy.
Add instructions for one repository
Repository settings combine with the global policy. Before you save, use the preview to examine the final instructions and policy.
A repository override can make the combined policy more permissive. Preview the merged result. Test the merged result.
Preview combined instructions
Use the prompt preview to examine the instruction order. Check the administrator instructions, global rules, repository rules, and added guidance.
The preview does not include session-dependent tools, skills, or runtime context. Use a test session to make sure that those parts operate correctly.
Set approval policy
Each action category can be:
allow— run without a person approving itrequire_approval— pause and show the full actiondeny— do not offer or run the action
Categories include read operations, external MCP reads, workspace writes, shell execution, Git writes, network access, Kubernetes mutations, and external MCP writes. Allow read operations first. Require approval for consequential actions. Deny actions that are not necessary for users.
Capabilities can be available to users, reserved for administrators or the system, or disabled. You can also limit custom Agent creation.
Configure MCP and workspace backends
- Use MCP integration to add shared tools.
- Let users complete their own connection fields.
- Use Workspace backends to select where new workspace-backed sessions run.
After a backend change, start a test session and verify a new workspace.
Configuration API
Use Settings for interactive administration. For automation, use the authenticated v2 routes in the deployment API reference:
GET,PUT, orPATCH /api/v2/ai/agent/runtime-config- repository runtime configuration below
/api/v2/ai/agent/runtime-config/repos/ - instruction rules, templates, and prompt preview below
/api/v2/ai/admin/agent/
See API overview for authentication and discovery.
Change check
- Preview the combined global and repository instructions.
- Start a new Agent session as a representative user.
- Make sure that the intended model and tools are available.
- Do one read, one approval-required action, and one denied action.
- Before you expand access, examine the session audit.