Claude Code Tip: Conserving Context

Here's a quick tip for anyone using Claude Code: if you want to conserve your context window and avoid auto-compact, ensure everything runs in a sub-agent.

Why?

For several months now, I've been using Claude Code to write 100% of my code. Recently I've been working on a project with huge files and long command output. I found that working in this project brought Claude to a halt because of constant auto-compaction.

After complaining about this to a colleague, they reminded me that sub-agents have their own context window.

Context preservation

Each subagent operates in its own context, preventing pollution of the main conversation and keeping it focused on high-level objectives.

Now, I had already been using sub-agents for all code writing and complex debugging tasks. However, I had a ton of other "routine" tasks that were not using sub-agents. This included reading files during the planning phase and executing git operations once work completed.

So, I immediately did a few things that solved the issue.

Sub-Agents

I created three sub-agents that were tasked at running certain tasks and providing a summarized output.

1. Created a code-researcher sub-agent.
2. Created a command-runner sub-agent.
3. Created a specialized git-expert sub-agent.

They are clearly instructed to provide smaller outputs to prevent the main context window from filling.

Instructions

I updated instructions in various places to ensure Claude Code knows to use these sub-agents.

1. Instructed CLAUDE.md to use code-researcher for all research/file reading tasks, including any time I ask it a question or to make a plan.
2. Instructed CLAUDE.md to use git-expert to run all git commands.
3. Instructed CLAUDE.md to use command-runner for any command that does not have a more specific sub-agent to run it with.
4. Instructed my slash commands like /ship, /push, /git-merge etc. to use the git-expert sub-agent for all tasks.

Results

As seen in the screenshot above, I am now able to execute much longer workflows without dealing with auto-compacting. Overall, this seems to be leading to shorter task completion times and more consistent output from Claude.

Previously, the above workflow would have required a clean context or would have caused an auto-compact.

Experiment

(Warning: this experiment could cost you money. Proceed carefully.)

To test this approach, identify a routine that fills your context window and causes an auto-compact. For example, use an MCP server that uses many tokens, like the Atlassian JIRA server. You could repeatedly ask it to give you the details of your current tickets.

First, run that routine with Claude instructed not to use sub-agents for the tasks. See how far you get.

Then, run that routine with Claude instructed to use your newly-created @agent-command-runner. You should be able to run signficantly more commands before filling your main agent's context window.

Overall Token Usage

An important caveat: your overall token usage will increase.

For example, a routine that uses 150k tokens without sub-agents may use 250k tokens with sub-agents. This is because Claude Code uses tokens to communicate between models. So, if you are operating in a cost-sensetive environment, such as on a low-price Anthropic plan, you may want to be careful with this approach. While you will experience fewer auto-compacts, you could hit your plan's usage limit faster. If you use Amazon Bedrock or similar, it could lead to higher bills.