What happened
Aravind Srinivas shared research on Perplexity Computer's Brain, which uses past session trajectories to propose personalized agent skills when it finds recurring patterns with enough evidence. In an eight-week randomized test on production traffic, 72.9% of users received at least one new personal skill, compared with 6.5% who added one manually. Tool call failures fell 11% and dissatisfied threads fell 7.6%, with no measurable change in cost or latency.
Why it matters
An eight-week randomized test on production traffic found fewer tool call failures and dissatisfied threads, while cost and latency showed no measurable change; shared skills also undergo replay testing and engineer review.
Sharing new research on self-improving skills in Perplexity Computer. Brain, which already maintains a knowledge wiki about your work, now also writes and updates agent skills for you based on your past sessions. An improvement agent reads the full session trajectories (tool calls, reasoning, results), finds recurring patterns, and only proposes a new skill or an update when there’s enough evidence. Built-in skills that every user gets go through paired replay on failure scenarios and engineer review before any change goes live. In an 8-week randomized test on production traffic, 72.9% of users got at least one new personal skill vs 6.5% who added one manually. Tool call failures dropped 11% and dissatisfied threads dropped 7.6%, with no measurable change in cost or latency.
View the quoted post on XComputer's personalized self-improvement agent, Brain, now automatically creates self-improving agent skills for you.
Brain learns how you work from past sessions and memory, then writes and updates personalized skills for future tasks.
The story so far
Perplexity said Brain, its personalized self-improvement agent in Perplexity Computer, now automatically creates and updates personalized agent skills for future tasks. According to Perplexity, Brain learns how a user works from past sessions and memory, and proposes new skills when it finds recurring patterns with enough evidence in past session trajectories. Aravind Srinivas shared results from an eight-week randomized test on production traffic evaluating the feature. In the test, 72.9% of users received at least one new personal skill, compared with 6.5% who added one manually, Srinivas said. Tool call failures dropped 11% and dissatisfied threads fell 7.6%, with no measurable change in cost or latency.
