Curated topic · 14 ARTICLES
Explore practical AI Agents articles grounded in clear explanations and engineering tradeoffs.
Latest articles
A harness is the execution loop wrapped around a model. We cover its components—tools, permissions, context management, and feedback—the reason agent performance is determined by the product of model intelligence and harness quality, and what makes a good harness.
Read articleA larger context window is not inherently better; higher relevance is. This article summarizes four strategies—selective loading, summarization, isolation, and externalization—for using a finite window like a budget, along with design criteria for always-loaded files such as CLAUDE.md.
Read articleThe AI covered in the previous seven installments was ultimately a “ask-and-answer” tool. People drove the conversation: refining questions, providing materials, and verifying answers. But the next step the AI industry has pursued for the past two years is different in kind. AI agents do not just converse—they take on the work itself.…
Read articleWhen using AI coding tools, you often repeat the same instructions: “Use this format for commit messages” or “Follow this checklist before deployment.”
Read articleIn earlier installments, we explored ways to conserve context: clear history at each work boundary (Part 3) and keep fixed overhead short (Part 4). Yet some tasks remain too large to handle. A typical example is, “Figure out how payment logic flows through this codebase.” A diligent agent would read dozens of files, and as we saw in Part 1, everything it reads accumulates in context. By the time the investigation is complete, no context remains for the actual work that knowledge enables—changing the code.
Read articleIf you have followed the series this far, you will see a common pattern: minimize what stays inside the context window. Part 3’s “write state to a file before /clear,” Part 4’s “only pointers instead of the body,” and Part 5’s “leave the process to subagents, send only conclusions to the main agent” all point in the same direction. But we have not yet properly addressed the destination of these files—that is, information sent outside the context.
Read articleIn Part 3, we saw how to clear conversation history with /clear. But checking the remaining context immediately afterward shows it is not at 100%. Run /context in Claude Code and the reason becomes clear: the system prompt, tool definitions, CLAUDE.md, and tools registered by MCP servers already consume tens of thousands of tokens…
Read articleWhen using a coding agent like Claude Code, you may see a warning such as "Context left until auto-compact: 8%" at the bottom of the screen. Chat-based AI does not usually show this. Yet surprisingly few people can explain why the number decreases or what happens when it reaches zero…
Read articleIn Part 1, we saw why the context window is finite. But looking at today's model specs raises a question: models with 1 million tokens already exist, so why not just feed in the entire codebase or document? Is there any point in being careful?
Read articleThe previous two posts completed the diagnosis. The context window is finite (Part 1), and filling it does not mean everything is used; performance actually declines as it grows (Part 2). The prescription is clear: manage context. Coding agents, including Claude Code, provide /clear and /compact for this purpose…
Read articleIn my previous comparison of six macOS terminals, I said I would introduce a separate terminal for the age of AI agents. This is that follow-up.
Read articleThere are more ways than ever to extend AI coding tools: connect tools with MCP, register procedures with Skills, and delegate work to subagents.
Read articleWhen using Claude Code in a project, two needs arise: “Always follow this rule” and “Remember what we discovered last time.”
Read articleUsing ChatGPT or Claude can leave you wanting more. The models are smart, but they cannot see your database or read your internal wiki.
Read article