An agent can act, not just suggest
An enabled agent workflow can run commands or work with project files in an authorized environment. That is more authority than receiving a text suggestion. Use a dedicated, appropriately permissioned account and a project whose contents you are prepared to expose to the selected tool.
Project file and Git automation requires explicit project permissions and the selected host's working confinement backend. Supported backends include Linux bubblewrap, macOS broker/Seatbelt and Windows AppContainer; availability depends on the installed agent and its capability checks. Use an enrolled compatible host when the required backend is unavailable locally.
Authorized project tools can create isolated worktrees and fork sessions when the runtime offers those capabilities and the owner has granted the required project and administrative permissions. A raw repository path or an ordinary terminal launch does not grant this authority. Check the available tools for the selected project before requesting a fork or worktree.
Prepare and supervise a task
- Select the intended agent, server and project.
- Review AI tools on this server where available, including compatibility and credential status.
- Refresh the tool inventory when the environment has changed. Installed does not necessarily mean ready or authorized.
- Review the requested task, permissions and any spending controls before starting it.
- Follow activity and results in the enabled task view.
- Inspect the resulting files, commands or remote state before accepting the work as complete.
A task status is not a substitute for verifying the outcome. Keep backups and review changes before relying on generated work in production.
In AI Tasks, choose New task, write the instruction and use Save task to save a draft; saving does not start an agent. Where the multi-host orchestrator is available, choose Confirm Plan Before Dispatch to review its plan. With a connected host and its installed coding tool, Session Action → Talk to Existing Session sends a follow-up to a selected existing session, while Start New Session creates a separate session. Review the host, session and working directory before dispatching.
Budgets and cancellation
Open a saved task's Objective budget controls where available. In Recorded spending, distinguish the configured limit, reserved amounts, recorded spend and remaining allocation. Pending reservations are unavailable to spend. Save limit updates the objective's USD allocation; zero does not mean unlimited. Use Refresh spending to check an unconfirmed save or an older snapshot.
Strict limits requires a bounded cost within the remaining allocation before server admission. Best effort, and its optional Allow tasks with unknown cost, can permit charges above the estimate or budget. Delayed provider reporting and unknown-cost attempts are not proof of zero cost. A displayed budget is not a universal cap on every third-party bill.
Account allowance and credits shows a separate account snapshot for MCP calls, managed relay usage and prepaid credits. Those credits are separate from the objective's USD allocation and the AI provider's charges. Missing quota or pricing data means unknown, not free; use Refresh account costs to check current information.
Stopping a task or losing the app connection may not undo a completed command or terminate every detached process. Confirm remote state for sensitive operations. A supported isolation mechanism also does not make all permitted file contents or commands harmless.
Use connected assistants to review external tools with persistent access, and privacy and trust before granting access to private projects.