Harden the System Prompts in AI Workflows with Agentic Radar

Previously, Agentic Radar focused on making your agentic workflows transparent and detecting potential vulnerabilities. It could map out agents, tools, MCP servers, and their dependencies.

Now it goes one step further. Run a scan on your workflow, and Agentic Radar will:

• Extract system prompts for each agent it finds

• Analyze them for ambiguity, vagueness, or risk factors

• Suggest hardened alternatives that reduce risk and improve performance

Think of it as static analysis for your agent’s brain.

In this section, we’ll walk through a real-world example where Agentic Radar scans an agentic workflow, identifies a weak system prompt, and suggests targeted improvements. We’ll keep things hands-on, so feel free to follow along with the code and try it yourself.

Initial Setup

1. Install Agentic Radar by following the README instructions from the official repository.

2. Download (or copy/paste) the agentic workflow written using OpenAI Agents from here.

3. In your terminal, set your OPENAI_API_KEY environment variable by running: export OPENAI_API_KEY=<your-openai-api-key>

Scanning Agent Vulnerabilities and Mitigations

Let’s scan the example workflow with Agentic Radar:
agentic-radar scan openai-agents -i path/to/example/folder -o report.html

Afterwards, open the generated HTML report in your browser.

In the example we used, we have two agents:

• Customer support agent - this agent does the actual job

• Guardrail agent - this agent helps the customer support agent by ensuring it doesn’t allow off-topic conversations (in this case, solving math homework for the user)

Agentic Radar performs static code analysis to detect all agents within the workflow. In addition, it also recognizes which mitigations are already implemented for some of the most common LLM vulnerabilities.

You can see that for certain vulnerabilities (Input Length Limit and Intentional Misuse) there are already some mitigations implemented in the code.

Hardening the System Prompt

Agentic Radar also shows us the currently used system prompt for the customer support agent.

While the system prompt is short and simple, it's also too vague and can lead to ambiguous behavior, inconsistent responses, or even unintended model actions. Without clear boundaries or task definitions, the agent may interpret instructions too broadly.

Now, let’s run the same command, but now including the –harden-prompts flag:

agentic-radar scan openai-agents —harden-prompts -i path/to/example/folder -o report.html

After opening the new report, you should see the suggested hardened prompts alongside the current system prompts.

You can click on the copy button (on the right) and replace the old system prompt in the source code with its improved version. You can do that by passing the copied prompt to the instructions parameter of the Customer Support Agent constructor.

Let’s scan the workflow with Agentic Radar again, now with the hardened system prompt:

agentic-radar scan openai-agents -i path/to/example/folder -o report.html

Now let's take a look at the Agent Vulnerability Mitigations section again.

Agentic Radar immediately recognizes that we have employed partial mitigations when it comes to PII Leakage and Jailbreak, as we added a layer of defense through our hardened system prompt.