AI Agent Architectures — What Your Team Is Building

My engineering team is building an AI agent for spec-driven development. Six months ago, I had no idea what an "agent" was in the AI context. If you are a PM managing AI-enabled teams, you need to understand the basics.

What Is an AI Agent

An AI agent is a system that uses a large language model to make decisions and take actions autonomously. Unlike a simple chatbot that responds to prompts, an agent can break a task into steps, use tools, call APIs, and iterate on its own output. Think of it as the difference between asking someone a question and asking someone to complete a project.

The Frameworks

Three frameworks dominate right now. LangChain is the most mature and offers the most flexibility. CrewAI focuses on multi-agent collaboration where different agents play different roles. AutoGen, from Microsoft, emphasizes conversational patterns between agents. Each has tradeoffs in complexity, reliability, and maintainability.

What PMs Need to Know

Agent projects are fundamentally different from traditional software. The output is non-deterministic, meaning the same input can produce different results. Testing is harder because you cannot write simple unit tests for decision-making. Costs scale with usage because every LLM call has a price. Latency is higher because agents often make multiple LLM calls per task.

How I Manage Agent Projects

I track three metrics that traditional projects do not need. First, cost per task — how much does each agent execution cost in LLM API fees. Second, task completion rate — what percentage of tasks does the agent complete successfully without human intervention. Third, hallucination rate — how often does the agent produce incorrect or fabricated output.

These metrics give me a clearer picture than velocity or story points ever could for AI agent work. If you are managing these projects, start measuring what actually matters.