A note on Context Windows

In long agent conversations, Developers often lose track of how much context has been used. Tools like Claude Code or Codex have auto-compaction built in, but they compact too late. Opus 4.8 has a million token context window, but retrieval quality falls off after about 200k. Opus 4.7 was actually a huge regression in this, 4.8 has a lower drop off. The bigger effect though is the cost. This comes directly from the horse's mouth, longer context == more cost.

A few things that have worked for me:

• Use /context or /status to see where you currently are.

• Use /statusline to add context and more information to the bottom of your terminal

• /compact often, in both Codex and Claude Code

• Start new conversations frequently (like one per bugfix). Use the /new or /clear command

• Keep a small AGENTS.md/CLAUDE.md file. Avoid repeating instructions.

• Use fewer and smaller MCPs, skills and subagents. Every skill and subagent you have are injected into the initial prompt, inflating context. MCPs are loaded on-demand, but descriptions are still sent for the model to decide. By 3-4 turns you're already exceeding the optimal window. I have seen people use 20+ MCPs! Don't do that. Think about what needs to be scoped to the user vs the project.

• Spawn subagents. They give a small, concise results to your main agent, preventing context inflation. Also better retrieval! Most agents spawn them without your intervention.