Plans & Approval Workflows

This guide walks you through reviewing and approving plans — structured proposals your agent creates before tackling complex work.

When you ask for something with multiple steps — research, analysis, building several artifacts — the agent proposes a plan first. You review it, adjust anything that's off, and give the green light before work begins.

When the agent proposes a plan

The agent creates a plan when the work is complex enough that starting without alignment would be risky — multi-step analysis, research across sources, building multiple artifacts. You don't request a plan. The agent decides when one is warranted. Simple requests just happen.

What a plan looks like

A plan appears as an interactive card in chat with:

• Title and objective — What the work is and what it accomplishes.

• Approach — The strategy the agent plans to follow. Your chance to correct course early.

• Objectives — Ordered milestones, each with optional sub-objectives that break down the work further.

• Permission gates — Actions that might need your sign-off, like sending messages or calling external APIs.

Exploring sub-objectives

Each step in the plan has an expand button on the right side of the row. Click it to see the sub-objectives nested under that step — a numbered list of the smaller pieces of work the agent plans to do within that milestone.

If you're reviewing a plan that hasn't been approved yet, expanding a step also reveals the annotation field for that step (see Leaving per-step feedback below). On an already-approved plan, expanding a step shows the sub-objectives only.

Approving a plan

Click Approve at the bottom of the plan card. Before you do, configure two things.

Objective gates

Each objective has a gate control:

• Don't stop — The agent works through without checking in (default).

• Stop before — The agent pauses before starting this objective.

• Stop after — The agent completes the objective, then pauses so you can review.

Mix and match. Let data gathering run freely, but require a check-in before the final report.

Permission gates

Each permission gate has three settings:

• Allow — The agent performs this action without asking.

• Review — The agent asks for approval first.

• Block — The agent skips this action entirely.

Once configured, click Approve and the agent starts immediately.

Watching a plan in progress

Once a plan is underway, the card updates in real time. Below each active step you'll see a live status line — a short note from the agent describing what it's working on right now. This updates as the agent moves through the work, so you always know where things stand without having to ask.

Modifying or redirecting a plan

If the plan is close but not right, type feedback instead of approving:

Skip the competitive analysis — focus on internal data only. Add a step to export results as CSV.

The agent revises and presents an updated card. Go back and forth until it's right — work doesn't start until you approve.

If the plan misses entirely, just say what you actually need. The agent adjusts — simpler plan, or no plan at all if the request is straightforward enough.

Leaving per-step feedback

Typing in chat works for broad redirections. For targeted feedback on specific steps, use the annotation controls directly on the card.

Annotation controls appear on top-level objectives only — sub-objectives don't have a separate + icon. Click the expand button on any top-level step to open it, then use the notes field to add feedback for that step. Add your note, then move to the next step you want to change. Once you've added a note, the icon becomes a pencil — click it again to edit.

As you add notes, two things happen:

1. The step highlights to confirm it has feedback attached.

2. The note count in the card header updates so you can see at a glance how many steps have input.

When you're done annotating, the Approve button at the bottom of the card changes to Request changes. Click it to send your feedback to the agent.

Obvious sends your per-step notes to the agent as a single message. The card enters a loading state ("Reworking plan…") while the agent revises. When the updated plan is ready, the card refreshes automatically.

Tip: You can annotate some steps and leave others untouched — the agent revises only what you flagged and leaves the rest in place.

Gate approvals during execution

When a plan hits a step you marked Stop before or Stop after, execution pauses and a gate approval card appears in chat. The card shows the task name, which step is waiting, and the type of approval required. Click Approve directly in chat to let the agent continue.

If you set a timeout on the gate and don't respond in time, the step follows whatever default behavior you configured when you approved the plan.

After approval

The agent works through objectives in order, pausing at any gates you set to Stop before or Stop after. For larger plans, the agent may split work across separate conversations running in parallel — results flow back to the main thread automatically.