AI
Agent Architecture and the Azure AI Agent Service
Azure AI Agent Service defines an agent as a microservice that takes unstructured messages, optionally calls other APIs, and returns messages. You can update an agent without rebuilding the whole system. The service runs on Azure AI Foundry, which combines models, tools, and deployment into one portal. Developers wire tools through a UI or a few lines of code, with function calling handled server-side rather than client-side.
How Saifr Models Financial Compliance as Two Specialized Tools
Saifr, incubated inside Fidelity Labs, built its Communication Compliance Agent on two models it published to the Azure AI Foundry model catalog in 2024. The first, Safer Retail Marketing Compliance, takes a paragraph and returns each non-compliant sentence with a risk level (low, medium, or high) and a violation category (promissory, misleading, exaggerated, or unwarranted) under FINRA 2210 rules. The second, Language Suggestion, takes a non-compliant sentence and returns a balanced replacement. GPT-4o mini orchestrates both as tools.
“So it’s a bit of magic really.” — Jono
The Compliance Agent in Action
Jono ran the compliance agent playground demo (06:12) with three test sentences: “I guarantee you’ll make lots of money,” “We’re confident that we can beat the S&P 500 index,” and “invest in Bitcoin to get rich fast.” The agent made four OpenAPI calls, visible in the trace. First it queried the retail marketing compliance tool and got all three sentences flagged non-compliant. Then it called the language suggestion tool once per sentence and returned a noticeably more measured paragraph.
Wiring Foundry Models as Agent Tools via OpenAPI
Each model deployed to the Azure AI Foundry catalog exposes a Swagger endpoint automatically. Jono ran the language suggestion REST query in Insomnia (12:47) and got three different compliant rewrites of “I guarantee you’ll make lots of money” on successive calls. The retail marketing compliance model returned risk levels and violation categories for four sentences in a single request. To wire either model as an agent tool, pull the OpenAPI 3.0 spec, paste it into the tool definition, and add one field: the servers entry pointing to your deployment URL.
Notable Quotes
So it’s a bit of magic really. Jono · ▶ 10:00
As you saw that it’s pretty straightforward to do it. Shubhi Raj · ▶ 15:32
I guarantee you’ll make lots of money. Jono · ▶ 6:26
Key Takeaways
- GPT-4o mini orchestrates two FINRA-trained models as OpenAPI tools to rewrite non-compliant financial copy.
- The retail marketing compliance model returns risk level and violation category for each flagged sentence.
- Wiring a Foundry model as an agent tool requires only its auto-generated OpenAPI spec plus a servers field.