AI BriefingDeep Agents v0.7 Released
Key point
Deep Agents v0.7 reduces default input tokens by 65% through context engineering optimization without sacrificing performance.
Details
Deep Agents v0.7 significantly simplified the Base Harness to maximize token efficiency and cost while maintaining model performance. The core of this update lies in context engineering that removes unnecessary context.
The key changes are as follows:
- Removed default system prompt: Deleted general guidelines and tool-usage-related text.
- Optimized tool descriptions: Reduced built-in tool descriptions by 43%.
- Made the Todo list optional: Excluded
TodoListMiddleware, which did not contribute significantly to performance improvement, from the default settings and made it optional.
Through these changes, base input tokens per basic agent turn dropped from about 6k to 2k, a 65% reduction. This leads to cost savings and speed improvements.
To verify performance, a new evaluation suite covering three categories—Autonomous, Conversational, and Long-context—was introduced. Testing on GPT-5.6-Luna, Gemini-3.6-Flash, Claude-Sonnet-4-6, and Claude-Opus-4-8 models showed that overall reward was maintained while tokens and costs decreased. Notably, for the Luna model, tokens decreased by 34% and cost by 15%, while performance actually improved by 4%.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.