Building AI Agents: A Practical Framework for Choosing the Right SDK

Every agent framework lists memory, tools, context management and guardrails. The feature lists read the same. What separates them is what each one decides for you: who runs the loop, what happens to a large tool result, where the agent is allowed to write, and how hard it is to switch model providers. Those decisions matter as much as the model. On ARC-AGI-3, the same frontier model scored 62.7% under one harness and 99.9% under another, and the better harness cost less to run. On Terminal-Bench 2.0, LangChain raised its Deep Agents score from 52.8% to 66.5% by changing only the harness.

In this hands-on workshop you build a minimal agent harness, watch it break on a long task, and measure what each missing piece is worth. Then you run the same task through five production harnesses (Claude Agent SDK, OpenAI Agents, Google ADK, LangChain Deep Agents and Pydantic AI) on the same model, and compare what each one did for you and what it cost you. You leave with a worksheet that takes your own agent to a decision (adopt, extend or build), and a repo you can point at your own tasks to test that decision before your team builds on it.

Learning Objectives

By the end of this workshop, you will be able to:

1. See what a harness does for an agent, by building one without it and watching where it falls short.
2. Measure harnesses yourself instead of trusting benchmarks, and know why one run proves nothing.
3. Compare five production harnesses by what they decide for you: who runs the loop, what you still have to build, and what you have to work around.
4. Reach a decision for your own agent (adopt, extend or build a harness) by ruling options out against your constraints.