Works, goals and plans

Give a task a clear finish line

A Work holds an objective and its completion criteria. Use the app menu’s Set a goal… action or the goal strip above the composer to create or edit one.

Describe an outcome and an acceptance check. “Improve the form” leaves the finish line vague. “Reject blank titles, keep valid titles unchanged, and pass the title-validation tests” gives both you and the agent something concrete to review.

The goal form separates Objective, Done when and a Checklist. Use these fields to make the outcome and verification explicit. Clear goal removes the goal record; it is not a stop control for a running provider turn.

The accepted goal belongs to you. The provider’s execution plan is tracked separately; updating that plan does not rewrite the goal you agreed to.

Follow the execution plan

Choose Plan in the workspace selector. Supported provider plan and task updates feed the execution-plan view and its completion chip. Quality signals can help you check whether reported validation covers the files that changed.

Use the plan to see what the agent says it is doing, then compare it with the transcript, diff and test results. A checked task list is not independent proof that the software behaves as intended.

Pause, resume and complete a goal

The goal strip provides the relevant pause, resume and completion actions. Pausing a goal and stopping a running turn are separate controls; use the turn’s Stop action when you need execution to halt.

Before marking work complete, review the acceptance criteria, current Changes view and actual validation output. A conversation can finish a turn while its wider Work remains incomplete.

Continuation and background activity depend on the provider. For example, Muse can report native goal and background-work updates. Do not assume a goal runs as an independent always-on service after the application or provider process exits.

Understand usage and budget stops

These are execution controls for their specific paths. They do not cap all usage across your provider account. Verified Claude and ChatGPT subscription routes omit these Helios API caps; API-equivalent dollar estimates are not bills. Equivalent dollar or token breakers are not established for Google, Muse or Kimi.

A budget stop preserves queued text and marks the Work as budget-limited. The app menu’s Clear budget block is an explicit decision to reopen it. Check your provider’s billing and quota before doing so.

Missions and agent activity

The Missions pane is an optional tandem integration. It lists missions from a separately installed engine and can be empty when that integration is absent. Ordinary Helios chats do not require it.

When the provider reports subagents, the Agent Dock shows their activity and input needs. Supported Claude subagents can be stopped individually. Use Background for reported work that continues between turns.