Chapter 2: Context Management in LLMs

At first glance, the progress of large language models seems easy to explain: bigger models, more data, more compute, better intelligence.

That is true, but it is only half the story.

Scaling laws help explain why model capability improves as scale grows. Some abilities even seem to emerge after a threshold: translation, reasoning, coding, planning, tool use.

But another evolution happened outside the model.

The way we use models improved.

From Prompting to Agents

At first, we relied on prompts. People studied how to assign roles, provide background, constrain format, add examples, and ask the model to reason step by step.

This was manual context management.

Then came retrieval: documents, databases, and knowledge bases were brought into the model’s context.

Then came tools: file access, code execution, browser use, APIs, workflows.

Then came MCP, skills, memory, project context, and agent systems.

Under the surface, all of these solve one problem:

How do we place the right information, goal, constraint, tool, and feedback into the system at the right time?

From Chatbot to Working System

A chatbot usually has a conversation.

A working agent needs more. It needs project context, task context, tool context, history, and evaluation. It needs to know where it is, what to do, what counts as done, and how to verify the result.

This is why tools like Claude Code and Codex feel different. They are not just better chatbots. They are models placed inside richer work contexts.

A capable model without context is like a smart person who just woke up and knows nothing about the situation.

A capable model with context is like a teammate who knows the project, has tools, understands the goal, and can check their work.

That is the key insight:

Better context management can change the practical capability of the same intelligent system.

This is why the idea matters for life.