Human Context in AI Era

In the AI era, people can use AI agents to do their work, especially in software development. There are many tools we can use to accelerate our work—Claude Code, Codex, and Cursor can handle much of the coding, and in many situations, they can even outperform humans. But if you actually try using agents to code, you’ll quickly find their limits. Agents are driven by LLMs; LLMs don’t have memories, and their responses depend heavily on the context window. If you don’t provide the right context, the results will likely be poor. That’s why context matters so much for agents.

But this post isn’t mainly about an LLM’s context; I want to talk about the human context behind these agents. Agents still follow human instructions, and they only do the right tasks when users share the right context. So humans need to manage their own context. We can work on multiple projects at once with agents, but it’s still difficult to switch contexts quickly and precisely.

AI coding isn’t just typing a few prompts into a text box and sending them to an agent at the moment. If you’re not “vibe coding” and you’re building real projects, you need clear standards and deliverables. That means paying attention to the coding process and test results. You end up acting like a project manager with a full view of the project: understanding requirements, UI design, system design, and API requests so you can direct the “engineers”—the agents.

In my daily workflow, I’ve found that agents take time to generate results and edit files. While they’re working, I often end up on Twitter or YouTube. When they finish, I switch back to check the results. This should save time, but in practice it doesn’t. The problem is that I then have to pull my mind back to work; my brain needs time to return to the table from the playground. To change this, I tried doing two or three tasks at the same time, but that creates new problems. Switching contexts in my mind is even harder than browsing the internet.

After reading others’ posts about using agents, I learned that we can run agents in parallel with the right setup. How do we do that? Maybe we should reduce project switching and divide work into separate time slices. Then we can plan larger tasks over the long run, handle small tasks with prompts, and stay focused on the current context. Have agents summarize their work after each task is done.