A Mind works through steps: it reads data, runs a query, changes an object, checks something or builds new work, and each step appears in the chat as a card. This page maps the families of steps behind each area of DataMind OS, explains how risky actions are handled, and shows what a failed step looks like. The cards and their controls are described in Tool cards and Tool History.
Each area runs its own family of steps. The exact steps depend on the Mind, your permissions and what you ask for, so read this as a map, not a fixed list for a chat.
| Where it runs | What it works on | Read more |
|---|---|---|
| A Warehouse Builder chat | Warehouse objects: facts, dimensions, views and transformations, checks and runs. | Warehouse tool cards in chat |
| A Lake Builder chat | The Lake: importing source tables, connectors, syncs and load settings; it can hand custom connector work to the connector's own Mind. | Warehouse Builder and Lake Builder |
| A connector's own conversation | One connector end to end: tap files and secrets for Connect Anything, or the recipe, streams, hosts and data access for Connect any API. | Connector operations and data access in chat |
| The Model Builder | A semantic model, round by round, reporting each round back into the chat that started it. | Model Builder |
| The Model AI panel | One semantic model: measures, columns, tables, relationships, folders, field parameters and tests. | Model AI |
| The Slides editor's AI panel | The deck you have open: drafting, reworking and checking it. | Write slides with AI |
| An ordinary Mind chat | Questions over the data the Mind can see: queries and charts, web search, knowledge search and skills, plans, and scheduling. | Query tools and saved queries, Knowledge search and skills in chat |
| Swarm children | Parts of one large job, spread across sub-agents that hand their results back. | Swarm agents |
Every step carries one of four grades, and the grade decides what happens before the step runs:
The grade belongs to the action, not to the chat. A step that serves more than one area keeps the strictest grade it has anywhere, so nothing becomes more casual when the same action runs from a different chat or panel. Within one card, a group of held actions is judged by its riskiest member, so one destructive action makes its whole card destructive; see Change approval in a builder chat. Every step is also checked against the person behind the chat: the Mind works as the person who messaged it, so it can change only what that person may change, and a step that needs a permission you do not hold is refused with the permission named in the reason. See Who can do what with Minds and Approvals, permissions and safety for AI actions.
Some steps only run in their own place, and the Mind says so instead of improvising:
Each area keeps its own conversation, so a request lands where its steps live.
For work that takes several steps, the Mind keeps a plan and updates it as it works. One step is in progress at a time, finished steps tick off, and a step that is skipped keeps the reason it was skipped. The Plan strip above the message box shows the plan while it runs, with each step's state, including Skipped; the strip and its updates are described in Tool cards and Tool History.
A step that does not go through is reported, never hidden:
It depends on the area rather than on one fixed list. Each area runs its own family of steps, and your permissions and your request decide which of them a particular answer uses. The table above maps the families, and every step an answer takes appears in the chat as a card.
Some steps only run in their own place: connector steps need a conversation attached to that connector, and deck steps need the deck open in the Slides editor with its AI panel. Builder chats and the Model AI panel keep their own scope rules, described on Approvals, permissions and safety for AI actions. Move to the right place and ask there; each area keeps its own conversation.
The grade of the action decides. Reads run as part of the answer. Ordinary changes follow the chat's acceptance mode. Destructive steps wait in Manual and Auto-accept, and consent steps wait in every mode. See Warehouse Builder and Lake Builder and Change approval in a builder chat.
No. The grade travels with the action: a step keeps the strictest level it has anywhere, so no chat or panel gets a looser version of the same step. See Change approval in a builder chat for how a held group of actions is shown.
The platform said no before the step ran, and the card reads Refused or Blocked with the reason. Common causes: a permission you do not hold, an object that has moved or no longer exists, work that is still in progress, or a rule the action breaks, such as deleting a connector folder that still contains connectors (move them out first). The Mind can propose another way. A refusal for a missing permission names the permission; the gates are described on Approvals, permissions and safety for AI actions.
It shows Error on its card, and the message behind it explains what went wrong; open the step to read it, in the chat or in Tool History. On Warehouse steps the details sit under Error details, with Copy error. Whatever finished before the failure keeps its result, and the Mind can try another way or ask you for what it needs.
Some work moves between chats. A builder chat can hand custom connector work to the connector's own Mind: the work continues in the connector's own thread, and its result arrives automatically as Report from the connector Mind. Nothing counts as finished before that report arrives, and the hand-off card shows its state: Working, Waiting for you, Finished, Failed or Stopped. See Change approval in a builder chat.
Yes, in a chat. Ask for the scheduling directly and it creates the task; if the idea is its own or comes from a hint, it asks you first. It needs your Add Task permission, and the new task inherits the chat's settings, such as the model and the Web, Compact and Swarm options. See Scheduled Tasks.
A skipped step is one the Mind decided not to do, often after a failure made it replan the rest; the plan keeps the reason next to the step. The Plan strip above the message box shows each step's state, including Skipped; see Tool cards and Tool History.