Skip to content
For Managed AI Platform users, see the cloud platform documentation
Agentic AI > Agent Assist > Workflows and prompting

Workflows and prompting

This page describes the interactive session: the welcome screen, the list models command, slash commands, and the three main workflows (Design, Code, Deploy), plus prompting tips.

Start Agent Assist

Run dr assist in an empty directory

Only run dr assist from a dedicated and empty directory. Running this command in a directory containing code or other files is unsafe. When you use the agent assist coding workflow, the assistant clones the DataRobot Agent Application Template repository into the current directory. This action can overwrite or conflict with existing files, damaging the existing project and degrading the accuracy of the assistant's output. Before running dr assist, if you're not in a dedicated directory, create one and open the terminal there (for example, mkdir my-agent && cd my-agent, then run dr assist).

To begin, run the dr assist command:

dr assist

When DataRobot Agent Assist is started, the welcome screen shows the welcome message, three options, and the help footer:

Output: Welcome screen
$ dr assist
🔌 Running plugin: assist
█████████
         █████             ███████                              ████████              ██
█████████                  ██    ███             ██             ██      ██            ██                    ██
              █████        ██      ██    █████  █████   █████   ██      ██   █████    ███████      █████   █████
██████████████             ██      ███       ██  ██         ██  █████████  ██     ██  ██     ██  ██     ██  ██
              █████        ██      ██   ███████  ██    ███████  ██   ██    ██     ██  ██     ██  ██     ██  ██
█████████                  ██    ████  ██    ██  ██   ██    ██  ██    ███  ██     ██  ██     ██  ██     ██  ██
         █████             ████████    ████████  ████ ████████  ██     ███   █████    ████████     █████    ████
█████████

Welcome! I help you design, code, and deploy AI agents.

What would you like to do?
  1. Design an AI agent     → Describe your idea
  2. Code an AI agent       → Load and implement an existing agent_spec.md
  3. Deploy an AI agent     → Deploy an implemented agent to DataRobot

Type /help for commands | /quit to exit

Input is collected at the $ prompt. Type a number (1, 2, or 3) to choose a category, or, describe the goal in natural language. If the first instruction is one of those numbers, the assistant treats it as referring to the predefined categories.

Built-in commands

Type list models at the $ prompt (no slash) to see the LLM Gateway model catalog, the model your session uses, and whether to switch models or continue. See the Environment and commands reference for details. Slash commands use the / prefix; for the full list and descriptions, see Slash commands. Typing / shows available slash commands and suggests /help for details. The /quit command exits the session.

Design an AI agent

Use this workflow when starting a new agent idea. The assistant asks clarifying questions about the agent’s purpose, capabilities, and requirements—including business or domain concerns (for example escalation rules, user scenarios, or tone) as well as technical details. It suggests tools and models, and produces an agent specification. The specification is displayed and written to agent_spec.md in the current directory (YAML format). A complete spec typically captures what you need to implement and review before code exists: the agent’s behavior and prompts, model choice, tool definitions, and how authentication and integrations should work, so both developers and stakeholders can align on intent. For field definitions and examples, see the Agent specification reference.

The specification can be refined over several iterations. The assistant limits clarifying questions to about two rounds before proposing an initial draft, then iterates on that draft with you. When choosing a model, it recommends a capable default from your LLM Gateway catalog unless you specify cost or other constraints. Type list models at any time to browse what is available in your organization.

Before simulation or coding, the assistant asks whether the Agentic Starter template’s default chat UI is sufficient or you need a custom frontend (for example a dashboard, data visualization, or multi-page app). That choice is recorded in the frontend section of agent_spec.md. Once you have a spec, you can simulate the agent to validate the design before coding.

Example conversation: Design a customer support agent
$ I want to create a customer support agent for an e-commerce site
Agent Assist: I'll help you design that! I can ask some questions to understand your requirements better. What specific customer issues should this agent handle?
User: $ It should handle order status, returns, and product questions
DataRobot Agent Assist: Great! What data sources will it need access to?
User: $ Our order database and product catalog
DataRobot Agent Assist: Perfect. Should it escalate complex issues to humans?
User: $ Yes, escalate billing and refund requests
DataRobot Agent Assist: I'll create a specification for your customer support agent...
[Creates and displays agent_spec.md, saves to file]

Simulate an agent

After you have an agent_spec.md, you can simulate the agent before implementing it. Simulation is sometimes called a dress rehearsal: you try the design as an end user would, without writing implementation code. Simulation is for design validation only—no real tool code runs, and no external services are called.

  • What happens: The model (LLM) uses your spec to decide which tools to call and with what arguments. When the model requests a tool call, Agent Assist does not run your tool implementation or call DataRobot deployments, datasets, or other APIs. Instead, it uses the LLM to generate simulated return values that match the tool’s described inputs and outputs. You see which tools were selected, the arguments passed, and the simulated results—so you can judge whether the design makes sense.
  • What it’s for: Validate that the right tools exist, that input/output structures are reasonable, and that the selected model tends to call the right tools for the task—before you spend time implementing real tool code.
  • What it’s not: Simulation does not call real DataRobot deployments, read real datasets, or hit real APIs. To use live data and deployments, implement the agent and run it in the Code or Deploy workflow.

Ask the assistant to simulate your agent (for example, “Simulate my agent” or “Let me try the design”) to begin. During the session:

  • Act as the end user—Type messages the way a real user would (for example order-status questions, research queries, or edge cases you care about).
  • Try specific scenarios—Describe particular requests or failure modes to see whether the model picks the right tools and arguments.
  • Capture design notes—If something in the prompt, tools, or examples should change, tell the assistant; it can update agent_spec.md based on what you learn.
  • End and review—When you are done, ask the assistant to wrap up simulation. It can summarize how the agent performed and suggest concrete changes to the spec (system prompt, tools, model, or examples).

Type /quit to exit the session when you are finished.

Code an AI agent

Use this workflow when the agent_spec.md file already exists and is ready to implement. The assistant loads the specification and—if the template repository is not yet present—clones the DataRobot Agentic Starter repository into your working directory. It then follows the template’s AGENTS.md for setup and implementation. The assistant can run shell commands (with approval), edit files, and manage multi-step work with tools. It suggests testing locally and deploying when the implementation is ready.

Empty working directory

Agent Assist clones the template into the current directory. If the directory already contains files other than agent_spec.md, the assistant warns you before proceeding so existing work is not overwritten.

When preparing the template for the first time, the assistant typically:

  1. Verifies dependencies—Runs dr dependencies check so required tools (Git, Node, Task, Pulumi, and others) meet minimum versions. See Prerequisite tools.
  2. Asks for an agentic framework—You choose one of the supported frameworks: LangGraph, CrewAI, LlamaIndex, NeMo Agent Toolkit (NAT), or Base (minimal scaffold). The choice is saved in the template configuration and determines which example patterns the assistant follows.
  3. Initializes the project—Runs dr start (with your approval) to configure credentials, ports, and LLM settings for the Agentic Starter template.
  4. Implements from the spec—Adapts template code to match agent_spec.md (tools, prompts, and frontend requirements). Tools that need external API access use the authentication described in the spec; you configure secrets in the template as described in AGENTS.md.

Windows

Agent Assist runs on macOS and Linux. On Windows, use WSL or a DataRobot codespace for the Code workflow. See System requirements.

The Agentic Starter template is structured for production-style agent applications (for example, an MCP server surface for tools, guided configuration through dr start, and deployment-oriented tasks). For details on what the template includes out of the box, see the Agentic application template overview.

After implementation, test the agent locally using the commands in the template’s AGENTS.md. Run those commands in a separate terminal in the project directory. When you are ready for production, continue with Deploy an AI agent.

Example conversation: Add a tool to an existing specification
User: $ Load my existing agent specification and add a tool for inventory checking
Agent Assist: I found your agent_spec.md. I can see it's a customer support agent.
Let me add an inventory checking tool...
[Reads current spec and adds the requested tool]

Deploy an AI agent

Use this workflow when an implemented agent is ready to deploy to DataRobot. The assistant reads the deployment instructions from the template’s AGENTS.md and guides users through the steps. It does not run deployment commands; it indicates what to run (in a separate terminal). For interactive or long-running steps (such as dr start), the assistant uses the appropriate tools as defined in the codebase.

Example conversation: Deploy the agent
User: $ Help me deploy the agent I implemented
Agent Assist: I'll read the deployment section of AGENTS.md and walk you through the steps...
[Displays deployment commands and instructions]

Interact with the DataRobot CLI

DataRobot Agent Assist runs certain DataRobot CLI (dr) commands on your behalf and can run other shell commands (including additional dr commands) with your approval.

  • At startup: The assistant runs dr dependencies check to verify that required tools (Python, Node, Git, Task, Pulumi, etc.) are installed and meet minimum versions. See Prerequisites and installation.
  • Code workflow: When you use the assistant to clone and prepare the template (for example, "Code an AI agent" or "prepare to code"), it runs dr start in the template directory to initialize the DataRobot project. You are prompted to approve before the command runs.
  • Any workflow: The assistant can run shell commands—including dr subcommands—after you approve. It shows the command and description, then prompts for approval before executing.

For full DataRobot CLI usage (authorization, deployment, and other commands), see the DataRobot CLI documentation.

Prompting considerations

Consider the following guidelines to get better results from the assistant:

  • Be specific: "Create a customer support agent that handles order inquiries and escalates billing issues" rather than "Create an agent".
  • Provide context: "I'm building a travel booking agent and need help with the flight search API integration" rather than "Help me with my code".
  • Ask for explanations: "Explain why this code isn't working and show me how to fix it" rather than "Fix this".
  • Iterate: "This is good, but can you make it handle edge cases like cancelled orders?" rather than accepting the first solution.