Skip to main content

Copilot agent mode and Auto-apply

Agent mode is how the Copilot handles requests in new conversations. Conversations started before agent mode keep working the way they did. Conversations in agent mode show an Agent tag in the conversation list.

How Copilot agent mode handles a request​

In agent mode the Copilot works through your request step by step instead of picking one kind of change up front:

  • It checks first. Before changing anything it can look at your current flows, data points, Custom APIs, tickets, and recent conversations, so it doesn't ask you for things it can look up.
  • It works from real data. It can pick real orders from recent tickets and read records from your connected systems, so drafts are built and tested against data that exists (see below).
  • It combines changes. One request can produce several drafts, for example a new data point and the flow that uses it, or an FAQ plus a canned response.
  • It reads its own results. If a step needs something that doesn't exist yet (a sample order, a Custom API), the Copilot adds that to the plan instead of stopping.
  • It asks once. When something is unclear, it asks all its questions in one reply.

Copilot plans in agent mode​

For work with several steps, the Copilot keeps a plan and shows it under its reply as a checklist. Steps are marked Ready to review when a draft is waiting for you, and Done or Applied when finished. Save a draft's card and the Copilot continues with the next step. To change the plan, just say so ("skip the WhatsApp template", "only do the FAQ part"); there are no separate run or skip buttons.

A later step can only build on changes that are saved. If a step depends on a draft you haven't saved yet, the Copilot stops and asks you to save that card first.

Real orders and live data in agent mode​

When a change depends on what your data looks like, the Copilot looks at real records instead of guessing:

  • Recent tickets of a flow. For account admins, the Copilot can list recent tickets a flow handled, with their order, the values the conversation collected, and the Custom API calls made for them. It uses these as test inputs, for example a real order number to try a new Custom API with.
  • Live reads from connected systems. The Copilot can read from your integrations to see where a value lives, for example an order's metafields in Shopify (Shopify is read with GraphQL queries; other systems with GET requests). These reads never change anything: writes are refused before they are sent. A reply makes at most eight live reads.
  • It remembers what it found. Facts it discovers about your systems, such as which metafield holds the invoice copy, are saved to your API documentation notes, so later requests don't need to look again.

Sample values are never made up. If no real record is available, the Copilot asks you for one.

Auto-apply for Copilot changes​

The Auto-apply switch next to the message box lets the Copilot save changes without you clicking Save on each card. It is off by default and remembered in your browser.

With Auto-apply on, a change is applied only when all of these are true:

  • you asked for that exact change (not something the Copilot suggested on its own);
  • the Copilot is confident the draft matches what you asked for;
  • the draft has no open questions and passes the same checks as saving its card;
  • nothing changed in the target since the draft was made.

Anything else stays a card for you to review. A reply applies at most three changes.

These always stay cards, even with Auto-apply on: Custom APIs, integrations, WhatsApp templates and flows, and product search actions. Flow changes are applied to the flow's draft only; publishing them is still your step.

Some changes to the disposition form also stay cards, because undo couldn't fully reverse them: deleting fields, changing or turning off filtering on a field that is filterable now, and adding or changing an execute action. Sample data is applied only when you gave the order or customer to fetch, and fetching it never changes anything in the connected system.

When a sample stays an Add Sample Data card, the Copilot can't use it until it is saved: click Fetch Preview, check the record, then click Save Sample. With Auto-apply on, the button reads Fetch & Save and saves the sample as soon as the fetch succeeds. Cards that only check data points against a record keep the separate Save Sample step.

Undo a Copilot change​

Every change the Copilot applied, automatically or from a canned response or business prompt card, is listed under its reply as Applied changes with an Undo button. Undo restores the previous version. If someone changed the same item after the Copilot did, undo is refused so their change isn't lost; edit the item directly instead. Settings changes are undone from their own settings card. Undo also works from the Copilot MCP server.