PostHog Desktop tasks

Contents

A task is the unit of work in PostHog Desktop. Every time you ask the agent to fix a bug, add a feature, or investigate something, you're creating a task. Each task is tied to a specific repository and persists across restarts, handoffs between your computer and the cloud, and handoffs between people.

Task modes

Tasks run in one of three places. The mode determines where the agent works and how changes are isolated from the rest of your work.

ModeWhere it runsIsolationBest for
LocalYour computer, current branchNoneQuick, small changes
WorktreeYour computer, dedicated branch + working directoryPer-folder isolationLonger work, parallel tasks on the same repo
CloudPostHog-managed workspaceFully isolated from your computerLong-running tasks

Local tasks change your current branch directly. Worktrees and cloud runs isolate changes from it. To run several tasks in one grid, use Command Center.

Autonomy modes

Separate from where a task runs, each task has an autonomy mode that controls how much the agent does on its own before checking in with you. Switch modes from the selector in the chat window, even partway through a task.

The modes depend on the harness the task uses.

Claude Code:

ModeWhat the agent does
Plan ModePlans without executing tools, so you approve the approach before any change
DefaultMakes changes, but prompts before dangerous operations
Accept EditsAccepts file edits automatically
Auto ModeAccepts file edits and shell commands automatically
Bypass PermissionsAccepts every permission request. Requires the Harness setting

Codex:

ModeWhat the agent does
PlanInspects and proposes, and makes no changes
Read onlyInspects files but cannot modify them
AutoEdits the workspace, and asks before risky operations
Full accessApproves every operation. Requires the Harness setting

Claude Code tasks start in Plan Mode and Codex tasks start in Auto. Change the default in settings, or set it to reuse the last mode.

Plan mode

In Plan mode the agent explores the problem and asks clarifying questions before writing any code. It surfaces options you can pick from, and if none of them fit you can type your own answer in the field. Once you approve the plan, it moves on to making the changes. It's the safest way to start anything non-trivial, since you align on the approach before the agent touches your code.

Queue and steer follow-ups

You don't have to wait for the agent to finish before sending your next message. A toggle in the chat window controls what happens when you send a follow-up while the agent is still working:

  • Queue adds your message to the back of the line. The agent finishes what it's doing, then works through your queued follow-ups in order.
  • Steer interrupts the agent at its next turn and redirects it right away, instead of waiting for the whole task to finish.

Next steps

  • Learn how environments prepare a task before the agent starts.
  • Learn how cloud runs keep a task going without your computer.

Still have questions?

Was this page useful?