How to Design an AI Agent: Architectures, Protocols, and Technical Evaluation

We are pleased to announce the release of Chapter 2: How to Design an AI Agent from our forthcoming book Agentic AI in Law and Finance.

Agents Are Architecture

Agents are not magic; they are architecture. In high-stakes domains like law and finance, the difference between a reliable tool and a runaway process lies in design. This chapter focuses on the architectural principles required to allow agents to function as cognitive work systems analogous to, and in concert with, professional teams.

Ten Fundamental Questions

We organize our analysis around ten fundamental questions that shape an agent’s operational reality:

Input Mechanisms

1. Triggers — When does the agent know it has work to do?
2. Intent — How does it understand what is being requested?
3. Perception — How does it find things out about the world?
4. Memory — How does it remember what it has learned and done?

Execution Strategies

5. Planning — How does it break complex tasks into manageable steps?
6. Delegation — How does it work with other agents and systems?
7. Action — How does it make things happen in the world?

Safety Layers

8. Termination — How does it know when to stop?
9. Escalation — When and how does it ask for human help?
10. Governance — How do we design systems that can be governed?

Design Decisions and Tradeoffs

Behind each of these ten questions lies a design decision with real tradeoffs. These choices determine:

• What a system can do
• How reliably it performs
• How it fails when things go wrong

The chapter provides concrete guidance for each decision point, with examples drawn from legal and financial applications.

Architectural Literacy

Ultimately, this chapter argues that robust design requires architectural literacy—a necessary bridge between technical implementation and professional obligation. Lawyers, compliance officers, and financial professionals need to understand enough about agent architecture to:

• Evaluate vendor claims critically
• Specify requirements meaningfully
• Oversee deployed systems effectively

Read the Full Chapter

The complete 95-page chapter is available on SSRN. This is the most technically detailed chapter of the book, providing the foundation for informed discussions between technical and professional stakeholders.

Together with Chapter 1 (What is an Agent?) and Chapter 3 (How to Govern an Agent), this forms a complete framework for understanding, building, and deploying agentic AI in regulated industries.