Perplexity's Brain System Automatically Creates and Updates Agent Skills

·2026.11.18 09:00

Key point

An eight-week randomized test showed that automatic skill generation reduced tool call failures by 11% and user dissatisfaction by 7.6% for US Max subscribers.

Details

  • Perplexity extended its Brain system to automatically create and update agent skills, allowing the Computer agent to learn procedural memory from user session history without changing model weights.
  • In an eight-week randomized production test, 72.9% of users with automatic skill improvement enabled acquired at least one new personal skill, compared to 6.5% in the control group, and the treatment group loaded personal skills 3.6 times more often.
  • The system reduced tool call failure rates by 11% and user dissatisfaction by 7.6%, with no measurable change in cost or response time.
  • Between August 1 and November 14, Brain wrote over 80,000 personal skills for more than 20,000 users. Additionally, between September 4 and November 14, over 100,000 updates were applied to more than 30,000 existing personal skills.

Improvement Mechanism

  • A four-stage loop (Evidence, Reflection, Verification, Activation) identifies past sessions where skills fell short and proposes updates.
  • Improvement agents run in sandboxes using frontier models to analyze full session trajectories and propose changes, with a per-user cap of 10 automatically created skills to maintain a curated library.
  • Built-in skills undergo stricter verification, including paired replay against failure scenarios and internal human review before merging.

Availability and Scope

  • Personal skill improvement is available to all US Max subscribers and is on by default for eligible users.
  • The feature is currently limited to personal skills; expansion to user-authored, project-shared, and organization skills is planned for the future.

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.