Model
Model is empty (Inherit) until you type an id or pick one from the same provider. Empty uses the model on the AI tab. The API key stays on the AI tab. If that model is not available for this Instance, this step fails. Images and Reasoning are Inherit, On, or Off. Inherit follows the AI tab. On is unavailable when the selected model reports that it cannot do that. Under Generation, an empty field (Inherit) follows the AI tab. Temperature, Max output tokens, and Reasoning effort (Minimal, Low, Medium, High) override that tab for this Agent only. Model turns with tools limits how many times this Agent can answer by calling tools in one run. Several tools in one answer count as one turn. Leave it empty for 50. The highest value is 500. Your plan may run fewer turns than the number on the Agent. When the turns run out, the Agent answers once more without tools, from what it already found, and says what is not finished.State snapshot
State snapshot copies records from a collection into the system prompt at the start of each run. Add up to four. Each row has a collection, a limit from 1 to 20, an optional sort, and an optional filter. The filter is a JSON array, the same shape as a State query filter. Leave it empty to copy every record up to the limit. Sort field is a field in the record, such aspriority, and Sort direction is Ascending or Descending. Leave Sort field empty to copy the most recently changed records first.
The copy is data. It does not add instructions, and values in the records are not filled in as templates. A tool call later in the same run does not refresh it. If nothing matches, the snapshot says No records matched. If the records cannot be read, the snapshot tells the model it did not load, so the model does not take it as empty, and the turn’s activity names the collection.
The inspector shows Name at the top. Settings holds prompts, the model, and Tools.
Tools on the card
Add Tools with + on the Agent card. The dialog is Add a tool: Search tools, then the catalog. With an empty search, Frequent lists up to eight Tools you have selected in this browser, most used first. Removing a Tool from the Agent does not change that list. You can add the same capability more than once with different settings. Two Tools on one Agent cannot share a name. Canvas Tool nodes are separate Graph steps. They are not required for the model to call a Tool. Another Agent as a Tool: open that Agent node to edit its prompts and Tools. The X on the Tool chip drops it. Back to agent returns from a Tool’s settings to the Agent. Call on that Tool sets how it runs. Wait for the result returns the other Agent’s answer in the same turn. Run in background starts a background task, and the turn goes on without the answer. The answer arrives in a later turn of the same chat, together with the task and its files. The calling Agent can pass up to 10 files by id; the background Agent gets them as files of its own message. Queued and running tasks are under Background tasks on Runs. On an Agent Tool, Agent decides lets the model set the value when it calls the Tool. Fixed always uses your value; the model cannot change it. Graph Tool nodes have no mode toggle — the field is always your value. With Agent decides, a choice outside its list or a number outside its range is refused: the Tool does not run, and the model gets an error with the allowed values or range so it can call again.Call limits per run
Open a Tool on the Agent card to see Call limits per run. Each action of that Tool has its own field: the actions you turned on for Telegram account, State, Files, Memory, Web, or Settings; the Tool itself for other Tools and for another Agent; each allowed tool for MCP. Enter a number from 1 to 100 to stop the model after that many calls in one run. Leave the field empty for no limit. Past the limit, the Tool is not called and the model gets an error. A limit counts only calls through this Tool, even if another Tool has an action with the same name. Turning an action off removes its limit. The same tool call with the same arguments runs at most twice in one run. The third time, the tool is not called. The model is told to use the earlier result or change the arguments. A different cursor, query, or page is a different call. Insert value adds Previous nodes or a Secret. Connect an incoming node first. Empty User prompt uses previous Result, not a named field such as Chat ID. Details: Previous nodes. Add secrets opens the account vault if the list is empty. Secrets still never go to the model.MCP
MCP (Remote server) is in the same Add a tool dialog. MCP.Emits
On the next node, Previous nodes listsAgent: Result ({{Agent.text}} while the name is unique). No extra named outputs.
Require approval before run exists only on an HTTP request node on the Graph. It is not available on HTTP tools attached to an Agent. See HTTP request.