Agent Mode

Agent Mode is Deskless’s multi-step coding workflow for tasks that need more context than a single selected block of code.

Availability and requirements

Cloud Agent Mode is available on iOS and Android. It requires Deskless Pro, an internet connection, and your own supported cloud provider API key: Claude, OpenAI, Gemini, Grok or OpenRouter. Provider usage is billed separately. See AI Setup to configure a provider.

New in 3.0 (Agent Mode 2.0): the agent edits a private draft, runs your project’s checks on the device before presenting its patch, saves checkpoints you can resume, and can work on its own Git branch. Deskless 3.0 is available for iPhone and iPad; Android 3.0 is in review on Google Play. The free Offline Agent is a separate Project Tools workflow using Apple Intelligence or Gemini Nano; it does not supply cloud Agent Mode access.

Start a task

1. Open the relevant project.
2. Open Agent Mode from the AI surface.
3. Describe the outcome, constraints, and files you already know are relevant.
4. Review the proposed plan before continuing.

Good requests name the intended behavior and important constraints without prescribing every line of the solution.

How the agent works

The agent progresses through a controlled loop:

1. Explore: read the relevant files and symbols in a snapshot of your project.
2. Edit a private draft: make changes in a draft that is separate from your files.
3. Check: run your configured tests and diagnostics with the bundled Project Tools, then iterate within a fixed budget of requests and tool calls.
4. Report: show what changed, which checks passed or failed, and the estimated provider usage.
5. Review: inspect every file and hunk, with suspicious edits and sensitive paths called out.
6. Apply: write only the changes you approve. Saving, committing and pushing remain separate actions.

A project needs a Project Tools configuration with tests for the agent to verify its work. Without tests, the report says so rather than claiming success.

Review controls

Approve or reject a whole file when the patch is straightforward, or review individual hunks when you want finer control. Rejected edits are not applied. An empty patch is treated as no change, not as a replacement file.

Task branches

Turn on Create a task branch before starting to have the agent work on a new agent/task-… branch from your current commit. It needs a clean repository and no unsaved drafts. Deskless never stashes your changes, overwrites an existing branch or switches back automatically.

Background work and checkpoints

Cloud requests and edits to the private draft can continue after you switch apps while the system allows it:

• iPhone and iPad: starting a task asks iOS for continued processing. When granted, the system shows the task’s progress and you can stop it there. Live Activities and Dynamic Island show agent progress on supported devices.
• Android: an ongoing Agent task running notification keeps the task going, with a Cancel task action. Android may stop it at its time limit.

Running tests, project code and language tools requires Deskless in the foreground; the task pauses before those steps until you return.

Every task saves checkpoints. If it is interrupted, open Saved Tasks to review it or choose Resume. Nothing resumes automatically, and a task resumes only in the same unchanged project, provider and model.

Safety checklist

Before applying a patch:

• inspect every file path and hunk;
• pay extra attention to build, security, credential, and deployment files;
• confirm the patch reflects the current editor buffer;
• run the project’s normal tests or validation after applying; and
• review the final Git diff before committing.

Agent Mode accelerates the work; it does not replace engineering judgment.