Tools and actor editor

Tools to be called and used by the agent and Actors to be assigned work are managed in separate editors. By creating tools and actors before connecting to an agent, multiple agents can use the same thing.

Tools

A tool refers to each function that an agent uses during work. On the list page, you can quickly find tools by type with the Type Filter Button and Count Badges.

type

Type	Description
Python	Custom tools written with Python scripts
Agent	Call other agents like tools

When you select the Agent type, the tool invokes another agent. The new-tool screen does not offer a Workflow type for calling a pipeline.

This procedure creates a separate tool asset. To select a delegate directly under Capability configuration → Agents, see Delegate to another agent.

Edit field

• Name / Alias / Description: Basic information. Use the AI generation button to draft a description.
• Type: Choose from Python / Agent.
• Input Schema (JSON Schema): Defines parameters to be passed when calling a tool.
• Execution Logic:
  • Python: Python script. In the Connector section you can choose which connection (connector) the script will use.
  • Agent: Agent to call + input mapping

Human approval is set on the actor, not the tool

The step of obtaining human approval before execution is specified by the actor's confirmation policy, not by the tool. Please refer to Confirmation Policy (HITL) below.

Run tests (Python tools)

The Implementation tab of the Python tool edit screen contains a Test Run section. Before saving the tool or connecting to an agent, you can run the code you're writing right away and see the results.

• Input Value: Enter the JSON object to be passed to the tool function. A default template is pre-populated with required parameters.
• Execution result: Success/failure, time taken, and return value (or error) are displayed as is. The results remain on the screen and are not separately saved as records.

Test execution actually runs the code

Since actual Python code runs in an isolated execution environment, external resources such as AI language models (LLMs), RAGs, and connectors can also be called. It is not a simulation that only mimics appearances. Test execution exists only in Python-type tools and does not appear in tools that reference agents/workflows or in built-in tools.

Configuration tab

Under Configuration for a Python tool, configure the execution environment and invocation guidance. An Agent-type tool shows invocation examples without execution-environment settings.

• Dependencies: Enter comma-separated Python packages, such as requests>=2.31. Whitespace and empty entries are removed when saved.
• Browser required: Enable this when the tool needs an environment that can browse and interact with web pages. Redeploy a running agent after changing its connected tools or this value.
• Tool invocation examples: Save multiple examples with a selection description and a required Arguments JSON object. A JSON array or string cannot be used.

Invocation examples are guidance saved in the tool definition to help an agent choose input. They differ from Test Run inputs, which remain only on the screen. Invocation examples appear for both Python and Agent types; dependencies and Browser required appear only for Python tools. All are read-only on a built-in tool.

Decorator bindings

Under Configuration → Decorator bindings for a Python tool or actor, connect runtime resources without editing source code. Enter the following for each row.

• Decorator: A registered name that begins with use_, such as use_connector, use_llm, or use_rag
• Binding name: A valid Python variable name used by the runtime function
• Arguments: An optional JSON object. Do not include the reserved key name; use the Binding name field instead.

The same decorator and binding-name pair cannot be duplicated. If a configuration row overlaps a source @use_* declaration with the same type and name, Configuration takes precedence appears and the configured value is applied. The source declaration is used when there is no configuration row.

The portal validates that the name uses the use_ format. A decorator that is not actually registered can be saved but fail when runtime resources are resolved. See Code to insert a decorator in source code.

Decorator bindings on built-in tools and actors are read-only. Use Copy and create from the list or editor to create an editable copy before configuring them.

Create by copying with built-in tools

Tools built in by the system include: The editing screen is read-only (displayed by the Built-in badge). If you want to change the logic or connected connectors of these tools to suit your environment, create an editable copy by clicking Copy and Create in the edit screen header. The copy opens with the original's name, collection, description, source, parameters, connector, and tags filled in, and all items can be freely modified before saving. The copy does not come with a built-in lock.

Built-in web tools

Tool	When to use it	Main scope
web_search	Find titles, URLs, and search snippets for relevant pages	Returns 1–30 results, with 10 by default. It does not return page bodies and treats no results as an error.
web_fetch	Read the static HTML body of one URL	Does not support PDFs, images, or JavaScript-rendered pages. The default body limit is 20,000 characters; truncated=true indicates truncation.
use_browser	Click, type, or scroll, or continue after static retrieval fails	Keeps browser state only within one call. The default is 12 steps and the maximum is 30.

In general, use web_search to find candidates and web_fetch to read a result URL. Switch to use_browser when extracted=false or the page is rendered with JavaScript. Browser results distinguish the conclusion in answer, extracted source text in extracts, and goal status in completed.

All three tools can restrict allowed domains and cannot access internal, loopback, or cloud metadata addresses. An agent connected to use_browser must be deployed with a browser runtime. There is currently no user interface for capturing a separate authenticated session.

Actors

An actor is each Python function that the agent calls and performs actual work. It is similar to a tool, except that it can require human approval (confirmation policy) before execution.

Edit field

• Name / Alias / Description: Basic information. Use the AI generation button to draft a description.
• Source code (Python): Write the def execute(**kwargs) -> Any function.
• Input Parameter (JSON Schema): Defines the input to be received when calling an actor.
• Connector: Select the connection (connector) the script will use.
• Confirmation policy: See below

Confirmation Policy (HITL)

In the Confirmation Policy section of the actor edit screen, determine whether this actor requires human approval before execution. Choose one of three.

policy	Action
auto	Run immediately without confirmation
Confirm	Requires human approval before execution (approval card shown in chat)
deny	Prevent execution altogether

When an agent attempts to execute an actor set to confirm, an approval request card appears in the conversation UI, allowing the user to approve or reject it. The approval flow and card UI are covered in detail in Agent Deployment Flow — Human Approval Card. The actor list provides filter buttons by verification policy.

Related UI

• Actor Selector: Used to search and select actors in the actor call node of the agent builder or AI agent settings.
• Actor Call Node Settings Panel: The panel that appears when you select an Actor Call node in the workflow agent allows you to edit mapping and execution options (including confirmation policies).

List screen common functions

• Name search and type filter buttons (Tool: Python / Agent, Actor: By verification policy)
• Each filter is given a badge with the number of items in it.
• Owner Column: Like other resource lists, the owner of each tool/actor is displayed.
• Select Edit / Duplicate / Delete in the action menu of the row.
  • Duplicate: Creates an editable copy with a new name and target collection. Both tools and actors can be cloned, even built-in items.
• If you try to delete a tool/actor that is referenced somewhere, you will now be warned in advance.
• Use Import in the list header and Export on a row to move definition JSON. See Import and export for file limits and agent bundles.
• To return to an earlier runtime configuration, open Version history and review the redeployment impact on referencing agents.

Next steps

• Create Agent
• Import and export
• Agent distribution flow