# Building the "Watchman" AI Agent

From Zero to Security Hero: Building a Wanted Persons Agent on Mastra & Integrating with [Telex.im](http://Telex.im)

### I. Introduction: Defining the Mission

* **Hook:** Briefly introduce the concept—using AI for security screening in chat platforms.
    
* **The Agent:** Introduce the **Watchman** agent and its purpose: checking names against the FBI Wanted Persons API.
    
* **The Tools:** State the core technologies: [Mastra.ai](http://Mastra.ai) (TypeScript), Gemini Model, FBI Wanted API, and [Telex.im](http://Telex.im).
    

### II. Step 1: Core Logic & Tool Development

* **API Choice:** Explain why the FBI Wanted API was chosen (free, public interest).
    
* **Mastra Tool:** Explain the creation of the `fbiSearchTool`.
    
    * *Key Success:* Defining the `inputSchema` (for the name) and the `outputSchema` (for the result) using `zod`.
        
    * *Code Highlight:* Show the `axios.get` call inside the `execute` function.
        
* **Critical Fix (403 Error):** Discuss the roadblock of the **403 Forbidden** error and how adding a `User-Agent` header to the `axios` call was essential for successful communication with the FBI API.
    

### III. Step 2: Agent Brain & LLM Configuration

* **Agent Setup:** Describe creating the `watchmanAgent` and defining its `instructions`.
    
* **API Key Crisis:** Detail the initial "API Key not found" error (`ECOMPROMISED`, `OPENAI_API_KEY` vs. `GOOGLE_GENERATIVE_AI_API_KEY`).
    
* **Solution:** Explain that you successfully switched the model to `google/gemini-2.5-flash` and used the correct `GOOGLE_GENERATIVE_AI_API_KEY` in the `.env` file, finally bringing the LLM online.
    
* **Instruction Focus:** Highlight the importance of the **disclaimer** ("not a guarantee of safety") and the **CAUTION** warning to meet safety requirements.
    

### IV. Step 3: The Integration Battle with Telex

* **Initial Setup:** Briefly mention using `npm run dev` and **ngrok** to get a public URL (`https://.../`).
    
* **The Great Protocol Mismatch (The Biggest Challenge):** Detail the persistent **404 Not Found** and **400 Bad Request** errors when using the standard A2A paths (`/api/a2a/agent/watchman`).
    
    * *Analysis:* Explain that the endpoint expects the *Chat Stream* format, but the Telex A2A Node sends the *JSON-RPC format*.
        
    * *The Workaround:* Explain that the final, successful integration required using the non-standard but functional `/agents/watchman/chat` path in the Telex Workflow JSON. This is where you overcame platform-specific routing challenges.
        
* **Workflow JSON:** Show a snippet of the final, working JSON (including your ngrok URL) to demonstrate the integration.
    

### V. Conclusion & Key Takeaways

* **Achievement:** The Watchman agent is fully deployed, functional, and successfully performs a high-stakes tool call.
    
* **Lesson Learned:** The biggest lesson was that backend integration requires **rigid adherence to protocol formats** (A2A vs. Chat Stream) and robust external **API error handling** (403 fix).