Deep Agent
Deep Agent is a model-agnostic AI agent component in RobilityFlow that can complete multi-step tasks without requiring users to design every action as a separate workflow step. Users define the objective, supply the required inputs, and make relevant capabilities available. Deep Agent uses its connected Language Model to decide what to do next, call the appropriate tools, and adapt its approach as results are returned.
For example, a user can upload monthly Excel sales reports and ask for a performance summary. Deep Agent can locate the files, process the figures, compare periods, and generate a report. The specific tools used depend on the request and the configured capabilities; file or script operations are not mandatory for every task.
How Deep Agent Fits into RobilityFlow
1. RobilityFlow passes the task instructions and available inputs to the Deep Agent component.
2. The connected Language Model interprets the objective and determines the next actions.
3. Deep Agent uses permitted built-in tools, Skills, connected component tools, or sub-agents as needed.
4. Tool results are evaluated to determine whether to continue, revise the approach, or produce the final response.
Unlike a fixed workflow, this execution path is not necessarily predetermined. The agent can make further tool calls based on intermediate results. For a step-by-step example, see How Deep Agent Works.
Capabilities and When They Are Used
Dynamic planning: Breaks an objective into actions and adjusts the next step when new results or errors arise.
Built-in and connected tools: Uses file tools, command execution, information retrieval, or capabilities exposed by connected RobilityFlow components when needed.
Skills: Follows reusable instructions and supporting resources, including SKILL.md-based guidance, for specialized tasks.
Workspace: Holds input files, working files, and generated outputs when a task involves file operations.
Terminal / Execute Tool: Runs commands or scripts when processing, calculation, or output generation requires execution.
Agent memory: Uses stored instructions or context when the relevant memory feature is enabled and configured.
Sub-agents: Uses the Task Tool to delegate separable subtasks and incorporate their results; delegation is optional.
Evaluation and approval: Reviews tool results and requests human approval for protected operations under the configured controls.
A simple response may not need a Workspace, Terminal, sub-agent, or file output. Deep Agent selects only the capabilities relevant to the task.
When to Use Deep Agent
Choose Deep Agent when the required actions cannot be fully defined in advance, the task involves several types of operations, or the next action depends on intermediate findings.
| Requirement | Suggested Approach |
|---|---|
| Repeat a known, fixed sequence of steps | Standard RobilityFlow activities |
| Analyze multiple files and decide how to process them | Deep Agent |
| Coordinate tools and adjust actions as results arrive | Deep Agent |
| Perform one predictable file or system action | Standard RobilityFlow activity |
| Generate a report from changing input data without designing each processing step | Deep Agent |
Example Scenarios
1. Document processing: Read multiple documents, extract relevant details, and prepare a consolidated report.
2. Research and analysis: Gather information from available sources and synthesize findings.
3. Data processing: Analyze datasets, transform records, and create structured outputs.
4. Script-based tasks: Create or modify scripts and execute them when permitted.
5. Workflow orchestration: Coordinate operations exposed by connected RobilityFlow components.
Execution Controls
1. Access to tools, Skills, and connected components depends on the agent configuration.
2. Workspace and command operations are subject to the configured execution environment and restrictions.
3. In the documented configuration, the Delete Tool requires approval before deleting files or directories.
4. Iteration, recursion, and context limits constrain execution.
5. If required information is unavailable or execution cannot complete, the response should identify the unresolved issue rather than imply success.
For operational details, refer to How Deep Agent Works, Workspace Management, and Skills Management.
