I'm wondering what the best practice is for using Warp to work on large projects. Am I better off starting with a new conversation for each discrete subtask, or to let Warp's summarization manage the context automatically. I know in Claude Code it is definitely the former. I want to keep context usage below 40% all of the time, if possible.
#New conversation vs summarization
shut dock
kind shadowBOT
Hi @shut dock. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:
The best practice in Warp is to start a new conversation for each discrete subtask, rather than relying solely on automatic summarization to manage context—especially if you want to keep context usage low (e.g., below 40%).
Here’s why:
- Relevance and Focus: Warp recommends starting a new conversation when your new question is unrelated to the previous one. Continuing in the same conversation with unrelated tasks can confuse the agent and bloat the context window, leading to degraded performance and less relevant answers. Starting fresh keeps the context focused and efficient Agent Conversations.
- Automatic Summarization: While Warp does automatically summarize earlier turns when the context window fills up, this process can sometimes lose important details, especially in long or complex workflows. Recent improvements have made summarization more robust (e.g., preserving TODOs and rules), but it’s still best to avoid overloading the context window when possible SWE-bench Verified Update.
- Manual Control: Warp provides visual indicators of context window usage, so you can monitor and proactively start a new conversation before hitting your threshold. This is especially important if you want to keep usage below 40% Context Window Management.
- Efficiency: Starting new conversations for new tasks also reduces unnecessary request usage and keeps the agent’s responses more targeted Warp AI Requests: Tip 2.
Summary:
For large projects, start a new conversation for each discrete subtask to keep context usage low and maintain high-quality, relevant responses. Use Warp’s context window indicator to monitor usage, and only rely on summarization as a fallback when you approach the context limit.