One Instruction, a Plan You Approve

From one sentence to a plan

Give an agent a goal in plain language and it doesn't just start acting on it. It breaks the goal into a sequence of steps, works out which steps depend on which others, and proposes the whole plan before doing anything.

Each step carries:

• what it accomplishes
• a hint at which tool it will likely use
• which earlier steps it depends on
• its own status: pending, in progress, completed, or failed

A step only starts once every step it depends on has actually completed; nothing runs out of order.

Approve, edit, or reject

The plan is shown to you before execution begins, not a checkbox you scroll past. You can:

• Approve it as proposed
• Edit it: reorder steps, or add and remove them, before anything runs
• Reject it outright

This is a different checkpoint than an agent's autonomy level (see Agent Management): autonomy controls how much an agent checks in during a run, while this is a single approval of the strategy before a run starts at all.

Watching it run

Once approved, each step runs as its own focused, bounded pass, capped at 10 iterations, so no single step can run away. You see steps complete one by one, in real time, with the currently active step highlighted, not just a final answer at the end.

If a step fails, it's marked failed rather than silently skipped; the agent can retry it, adapt, or hand the failure to you, depending on what actually went wrong.

Related

• Agent Management: autonomy levels, the related but distinct check-in behavior during a run
• Live Preview: the sandboxed environment a plan's steps actually run inside