Agents are very popular right now, which motivates many people to learn Agent development. However, it can be hard to know where to start. Based on my own experience, here are some thoughts. I hope they help.
Agent News and Resources
There is a lot of Agent-related content today. Much of it is traffic-driven content that mostly summarizes “second-hand” information from others. Some even make Agents sound like mysticism. These sources should be filtered out. Instead, focus on first-hand information from people with real development backgrounds and practical experience, such as the official Claude Code blog. Many of the posts there cover Agent technical concepts and directions and are excellent.
Frameworks vs. Context
When learning technology, many people start with frameworks—for example, learning Spring for Java development. There are also many Agent frameworks, but frameworks themselves often don’t lead you to the essence of Agents.
Interacting with a model is essentially sending a piece of text and receiving text back, which may be a result or a tool call. To get good responses from the model, you need to give it sufficiently reasonable prompt text for the current problem. This prompt text is essentially Context. Start with simple questions, then move on to analyzing the Context of mature Agents. There are already many mature Agents, such as Claude Code and Codex. You can analyze what they send to the model each time to gain a deeper understanding of Agent development. Frameworks should serve Context.
Get Hands-On
Many people don’t know what AI can do at first. The answer is simple: just start using it. Some people begin by having an Agent write git commits, and later find that letting the Agent commit and push directly works even better. Try using AI in different scenarios—this is the simplest and lowest-cost way to get started.
The same is true for Agent development. Don’t get lost in various Agent development concepts. As of 2026-08, Agents are still evolving rapidly and have not matured or stabilized. You must build Agents yourself, but don’t stop at demos and POCs—they must be usable in practice. The scenario can be very simple. If you can’t find a suitable scenario yourself, look at mature scenarios in the industry. For example, coding scenarios are relatively mature right now, and you can try building a coding Agent to learn.
Evaluate Your Agent
The biggest difference between models and traditional software engineering is that models are probabilistic and uncertain. You cannot guarantee that a certain approach will always produce deterministic results. Therefore, evaluation is essential. You don’t need complex evaluation frameworks or scoring. The simplest way is to use your own Agent. If your Agent generates presentations, try having it generate different kinds of presentations. Ideally, you should use it in your daily work rather than testing for the sake of testing. Good results mean good Agent development. If the results are not good, analyze the Agent’s execution trace to find which tool call or response deviated from your expectations, and then revise the relevant prompts. An Agent that can be put to practical use is a good Agent.
Self-Iteration
Traditional software development usually requires separate development and testing phases, but Agents can iterate on themselves. For example, the Hawa Code Agent is developed using Hawa Code itself. Developing a new feature becomes a validation and evaluation of existing functionality. This is one of the joys of developing a coding Agent. You can also try to achieve similar effects in other scenarios.