Deep Agent Component
Deep Agent processes complex tasks using connected Language Models, available Tools, and selected Skills. It interprets task instructions, determines the required actions, and performs them within an assigned workspace. This document describes its configuration parameters, execution behaviour, input handling, workspace structure, and outputs.
Prerequisites
1. Language Model: Connect a Language Model component.
2. Deep Agent Name: Specify a name for the Deep Agent instance.
Deep Agent Component Parameters
1. General
| Parameter | Description |
|---|---|
| Deep Agent Name | Specifies a unique name assigned to the Deep Agent instance to distinguish it from other instances. |
| Deep Agent ID | Displays the unique identifier generated automatically using the component ID. This field is read-only. |
| Language Model | Connects the Language Models used to interpret instructions, plan actions, and generate responses. Different Language Models can be connected for different tasks. |
| Tools | Specifies the optional Tools available to the agent during execution. Connected Tools and Component Tools can be configured for the agent. Refer to the Deep Agent Overview documentation for details. |
| Built-in Tools Requiring Approval | Allows selection of built-in Tools that require human approval before execution. When a selected Tool is called, execution pauses until an approval decision is received. |
| Auto Approve All Tools |
Determines whether Tool executions are automatically approved. Enabled: Automatically approves Tool executions without requiring individual approval. Disabled: Requests approval for Tools configured under Built-in Tools Requiring Approval. Default: Disabled. |
Approval Behaviour
a. Execution resumes after approval.
b. If rejected, the agent must not repeat the same operation with identical arguments.
c. The delete_file Tool always requires explicit approval, regardless of the approval configuration.
Note: The delete_file Tool always requires explicit human approval, even if it is removed from the Built-in Tools Requiring Approval list or Auto Approve All Tools is enabled.
2. Input Configuration
| Parameter | Description |
|---|---|
| Input Data | Accepts Message, Data, or DataFrame containing file_path values. These paths identify files to be copied into the agent workspace. |
| Input Files |
Allows files to be selected or uploaded directly. The files are copied into the FileInputs folder. Supported formats: PDF, TXT, CSV, JSON, MD, PY, DOCX, XLSX, PNG, JPG, and ZIP. |
| Input | Specifies the user message or task instructions that define what the agent must accomplish. |
| Agent Instructions |
Specifies custom instructions added before the base prompt to guide task execution and response generation. Default: You are a helpful assistant that can use tools to answer questions and perform tasks. |
Note: Input defines the task to be performed, whereas Agent Instructions provide additional guidelines for how the agent should perform the task and generate responses.
Execution Behaviour
a. Files selected or uploaded through Input Files and paths supplied through Input Data are copied into the FileInputs folder.
b. Duplicate file paths are removed before copying.
c. Existing contents of the FileInputs folder are cleared before copying the current input files.
d. Files with invalid or non-existent paths are not copied.
3. Chat History
| Parameter | Description |
|---|---|
| Number of Chat History Messages |
Specifies the number of previous conversation messages included as chat history. Set to 0 to disable chat history. Default: 100. |
| Context ID |
Specifies an optional identifier used to scope chat history to a particular conversation context. Default: Empty. |
Note: Increase the message count for follow-up tasks that depend on previous conversation details. Set it to 0 when chat history is not required. Context ID identifies the conversation context independently of the configured message count.
4. Workspace and Execution
Deep Agent uses the assigned Workspace and Terminal to perform file-related and command-line operations during task execution.
a. Workspace: Used to create, read, modify, search, and manage files.
b. Terminal: Used to execute scripts and command-line operations required to complete tasks.
Deep Agent uses both as needed, depending on the task requirements.
| Parameter | Description |
|---|---|
| Use Local Filesystem Sandbox |
Determines the filesystem backend used by Deep Agent. Enabled: Uses the local filesystem sandbox in the assigned workspace for the following operations: 1.LS – Lists files and directories. 2.Read File – Reads file contents. 3.Write File – Creates or writes files. 4.Edit File – Modifies existing files. 5.Delete – Deletes files. 6.Glob – Finds files using file patterns. 7.Grep – Searches file contents. 8.Execute – Runs shell commands and scripts. Disabled: Uses the alternative state-based filesystem backend instead of the local filesystem sandbox. Default: Enabled. |
| Max Iterations |
Specifies the maximum number of iterations the agent can perform to complete a task. Usage: Determines how many execution iterations are allowed. For example, a value of 2 allows the agent to perform up to two iterations. Default: 5. |
| Recursion Limit |
Specifies the maximum execution depth allowed for recursive operations. Usage: Adjust the limit to control recursion depth and prevent excessive recursive operations. Default: 500. |
| Verbose |
Determines whether additional execution and debugging information is logged. Enabled: Logs detailed execution information to help investigate and troubleshoot issues. Disabled: Does not log additional debugging information. Default: Disabled. |
5. Skills
| Parameter | Description |
|---|---|
| Enable Skills |
Determines whether selected Skills are made available to the agent during execution. Enabled: Copies the selected Skills into the agent workspace, making their predefined instructions and procedures available during execution. Disabled: Does not load or use the selected Skills. Default: Disabled. |
| Skills | Specifies the Skills selected from Skill Management for use during execution. This field is applicable only when Enable Skills is enabled. Refer to the Skill Management documentation for details. |
Execution Behaviour
a. Changes made to Skills copied into the agent workspace do not modify the original Skills in Skill Management.
b. The agent cannot directly access or modify the original Skill Management storage
6. Memory
| Parameter | Description |
|---|---|
| Enable AGENTS.md Memory |
Determines whether the agent loads project-specific instructions and stored memory from AGENTS.md files in the workspace. Enabled: Loads instructions, guidelines, and stored information from available AGENTS.md files to provide project context during execution and subsequent tasks. Disabled: Does not load or use instructions and memory stored in AGENTS.md files. Default: Disabled. |
| Memory Rules | Specifies what information the agent should store, retrieve, or update in memory. This helps maintain relevant context and reuse previously stored information across tasks and execution phases. |
7. Context Management
| Parameter | Description |
|---|---|
| Enable Context Summarization |
Determines whether the agent automatically summarizes conversation history when the configured token threshold is reached. Enabled: Automatically initiates summarization when the Trigger Tokens threshold is reached. The Keep Tokens threshold determines how much context is retained after summarization. This reduces the context passed to subsequent Language Model requests while preserving relevant information. Disabled: Prevents automatic summarization of conversation history. Default: Enabled. |
| Context Length |
Specifies the maximum number of tokens available for conversation context. This value is used to calculate when summarization starts and how much context is retained afterward. Usage: Adjust the value based on task complexity and the amount of conversation history the agent needs to retain. |
Context Summarization Calculation
The configured Context Length determines the following thresholds:
Trigger Tokens: Context Length × 0.85
Keep Tokens: Context Length × 0.10
Behaviour
a. When the conversation context reaches the Trigger Tokens threshold, summarization is initiated.
b. The Keep Tokens threshold determines the amount of context retained after summarization.
8. Rubric Validation
| Parameter | Description |
|---|---|
| Enable Rubric Validation |
Determines whether the response received by Deep Agent from the connected Language Model is evaluated against configured criteria before the final response is returned. Enabled: Evaluates the Language Model response against configured rubric criteria, such as accuracy, completeness, and formatting. Disabled: Skips rubric-based validation and returns the response without evaluating it against the configured criteria. Default: Disabled. |
9. Token Optimizer
| Parameter | Description |
|---|---|
| Enable Token Optimizer |
Determines whether large user prompts are compressed using the Token Optimizer before Deep Agent execution to reduce token usage. Enabled: Compresses large user prompts before execution to reduce token consumption while retaining relevant information. Disabled: Processes user prompts without Token Optimizer compression. Default: Disabled. Refer to the Token Optimizer documentation. |
| Token Optimizer Minimum Tokens |
Specifies the minimum estimated token count required for a user prompt to be considered for compression. Prompts below this threshold are processed without optimization. Default: 3,000. |
| Token Optimizer Target Ratio |
Specifies the target percentage of the user prompt to retain after compression. Usage: Controls how much context is retained after compression. For example, a target ratio of 50% aims to retain approximately half of the original context, reducing token usage while preserving relevant information. Default: 50%. |
Note: Token Optimizer Minimum Tokens and Token Optimizer Target Ratio are displayed only when the Token Optimizer is available and enabled.
10. File Management
| Parameter | Description |
|---|---|
| File Reuse Rules | Specifies instructions for identifying and reusing existing files during task execution. The agent checks for matching files, reuses or updates them where applicable, avoids unnecessary duplicates, and returns the resulting file paths. These rules help maintain file consistency and prevent unnecessary file creation. |
Output
| Output | Description |
|---|---|
| Response | Returns the final response generated by Deep Agent. Details of generated files are provided as execution metadata. |
Configuration Limitations
1. The Token Optimizer requires the Headroom service to be configured.
2. The Token Optimizer compresses user-provided context; it does not replace the entire conversation or system instructions.
3. File operations are restricted to the assigned workspace. The agent cannot access other agent workspaces or Skill Management storage.
4. Repeated Tool calls with identical arguments may be restricted by execution safeguards.
5. Increasing Max Iterations, Recursion Limit, or Context Length may increase execution time, resource usage, and token consumption.
