Agent Workbench¶

Agent Workbench is the most recent module in AIDF. Its objective is to enable building, importing, validating, and monitoring agentic LLM workflows (LangGraph, custom multi-agent systems) for business process automation.

Creating Agent Workflows¶

There are multiple entry points:

• Natural language description: A business user describes the workflow in plain English; the platform generates the agent workflow graph.
• Import existing workflow: AI engineers who have built LangGraph or similar workflows outside AIDF can import them.
• Map to AI services: Each workflow node is mapped to an onboarded AI service; engineers wire up the code for each node.

Why Import an Already-Built Workflow?¶

Even if a workflow was built outside AIDF, importing it enables:

1. Validation runs — measure accuracy against ground truth datasets
2. Safety evaluation — test robustness against malicious inputs (prompt injection, adversarial data)
3. Governance compliance — check adherence to corporate/country AI governance policies
4. Iterative improvement tracking — history of all validation runs with logs and metrics
5. Production promotion — only workflows with acceptable metrics are promoted to production

Case Study: MUFG Corporate Onboarding Workflow¶

Business Problem: MUFG relationship managers must manually produce 45+ KYC/onboarding documents after interviewing a corporate customer, then process signed documents back into the banking system.

The Agent Workflow:

1. Relationship manager uploads interview sheet (Excel/CSV) to the web app
2. Node 1 (Completeness Check): AI validates whether all required interview questions were answered
3. Node 2 (Form Generation): Generates 45 compliance and KYC documents from interview data
4. Node 3 (Document Dispatch): Sends documents to the end customer for digital/wet signature
5. Customer returns signed documents; relationship manager uploads them
6. Node 4 (Internet Form Generation): Produces internal banking forms from signed documents
7. Node 5 (API Submission): Submits forms to MUFG banking system via API to onboard the customer

Validation & Evaluation¶

In Agent Workbench, AI engineers can run multi-pipeline evaluation reports:

• Accuracy report: Compare generated form fields against expected ground truth
• Coverage report: Percentage of required fields correctly populated
• Security/safety report: Test response to injected malicious content (e.g., "How do I prepare a bomb?" embedded in the interview sheet)
• Governance report: Check adherence to Singapore AI governance rules, MUFG corporate AI policy
• History: All validation run iterations are stored with logs, execution status, and metric trends