QuestionQ88

VerifiedPrompt and Context Engineering

Your Claude application runs long agentic workflows in which the agent makes many tool calls, and the conversation history grows rapidly. After about 20 tool calls, you notice that the agent's responses become less focused and sometimes overlook earlier task constraints.

How would you address this?

Explanation

Tool-output pruning removes older results that are no longer needed, while compaction summarizes conversation history to retain relevant progress and constraints. Both techniques reduce context clutter while allowing the workflow to continue. Important constraints should be explicitly preserved during compaction. (platform.claude.com)

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!