명확함을 위한 5일: AI 에이전트 해명

Source: Dev.to

Overview

5일 AI 에이전트 집중 과정에 등록하기 전, 나는 에이전트에 대한 교과서적인 정의만 알고 있었다. 기본을 배울 것이라 기대했지만, 코스는 이론에서 바로 실습 랩으로 넘어가 코드를 실제로 작동시키는 경험을 제공했다. 다섯 번째 날이 끝날 무렵에는 에이전트를 배포하기 위한 모범 사례를 공부하고 있었다.

백서에서는 간단한 비유를 사용했다: 모델은 에이전트의 뇌, 도구는 손, 오케스트레이션 레이어는 신경계, 배포는 몸과 다리다. 이 비유는 ChatGPT에 프롬프트를 보낼 때 뒤에서 돌아가는 Think → Act → Observe 루프를 시각화하는 데 도움이 되었다. 또한 런타임에 새로운 도구나 에이전트를 생성할 수 있는 자기 진화형 에이전트 시스템이라는 개념도 소개했다.

Day 1 – Foundations

• Analogy: Model = brain, tools = hands, orchestration = nervous system, deployment = body/legs.
• Key concept: The “Think, Act, Observe” loop that powers an agent’s behavior.
• Insight: Agents can become self‑evolving systems, expanding their resources by generating new tools or agents on the fly.

Day 2 – Documentation & Integration

• Realization: Building AI agents isn’t just about technical know‑how; documentation and best‑practice protocols matter.
• Problem highlighted: The “N × M” integration challenge, where many agents and tools interact, can quickly become chaotic.
• Solution introduced: The MCP (Multi‑Component Protocol) to manage complex integrations.

Day 3 – Memory & Context Engineering

• Questions explored:
  1. If I tell the agent my favorite color is blue, will it remember that later?
  2. How does the agent update its knowledge when preferences change?
  3. Does it retain greetings like “Good morning”?
• Answer: An agent without memory is like an assistant with amnesia. Sessions and memory are essential building blocks for context engineering, regardless of the agent’s specialization.

Day 4 – Debugging & Evaluation

• Comparison: A calculator has a single correct answer (2 + 3 = 5), whereas a writer agent’s output is open‑ended.
• Challenge: Verifying correctness and tracing the agent’s reasoning process—did it call the right tools? Did those tools provide accurate information?
• Approach:
  • Implement LLM‑as‑a‑judge to automate evaluation.
  • Introduce a human‑in‑the‑loop for added reliability.
• Takeaway: Debugging an agent can be more complex and ongoing than building it.

Day 5 – Deployment & Multi‑Agent Communication

• Focus: Deploying agents and the A2A protocol, which enables different agents to “talk” to each other.
• Goal: Build agents that real‑world businesses can depend on, with continuous evaluation to maintain trustworthiness.
• Reality check: Fully trustworthy agents are still out of reach; human oversight remains essential.

Post‑Course Project

After the intensive, I spent the next 15 days building a project that applied all the learnings—from single‑agent design to multi‑agent systems, evaluation, and deployment.

Practical Example: SketchSensei

For anyone who has tried to draw a realistic human head and struggled with orientation and proportions, SketchSensei offers a solution. It overlays Loomis guidelines on an input image and generates step‑by‑step drawing instructions, letting you pick up a pencil and draw the head the Loomis way.

Acknowledgements

Thank you to Google × Kaggle for providing this course and equipping it with all the material needed to bring these concepts to beginners.