Perplexity says its Brain system gave 72.9% of users in a randomized test at least one new personal skill, against 6.5% of users without it who added one by hand, per its research blog. That is about 11 times the rate, by our arithmetic.
Brain already keeps a knowledge wiki about each user’s work. It now also writes and updates agent skills, the reusable instructions an agent loads for recurring tasks, based on past sessions.
Business Pill: The Cost of Checking
An evergreen explainer on who pays to check work, and why cheap verification changes what can be automated. It does not discuss Perplexity.
The key insight: Perplexity’s randomized test put automatic skill creation at 72.9% of users against 6.5% adding skills by hand, about 11 times the rate by our arithmetic, with tool call failures down 11%.
What the Randomized Test Found
Perplexity ran an eight-week randomized test on production traffic, comparing users with automatic skill improvement against a control group with it turned off, per the blog.
In the test group, 53% had a personal skill loaded in a conversation they started, against 13% of controls. Their turns loaded a personal skill about 3.6 times as often, the blog says.
Tool call failures fell 11% across all Computer activity, and about 7% when the system’s own background runs are excluded. Medium and high dissatisfaction threads per user fell 7.6%, with no measurable change in cost or response time, per the blog.

What Was Posted on X
Perplexity CEO Aravind Srinivas shared the research on self-improving skills: Brain writes and updates agent skills from past sessions, and in an 8-week randomized test 72.9% of users got a new personal skill versus 6.5% who added one manually, with tool call failures down 11%.
How Brain Decides What to Write
A daily improvement agent reads full session trajectories, including tool calls, reasoning and results, the blog says. It proposes a change rather than editing a skill directly.
Creating a skill needs two independent occurrences of the same behaviour, while one grounded occurrence can justify an update, per the blog. Each user is capped at 10 automatically created skills.
Built-in skills, which every user gets, change only after paired replay on failure scenarios and an engineer’s approval of a pull request, the blog says.
Scale So Far
Brain wrote over 80,000 personal skills for over 20,000 users between 1 August and 14 November, per the blog. That is about four skills per user, by our arithmetic from those lower bounds.
More than 100,000 automatic updates were applied to over 30,000 existing personal skills between 4 September and 14 November, the blog says. Software engineering was the largest skill category at 11.0%.
In an offline test on nine task families, task success rose from 43.8% without skills to 62.5% with generated skills, per the blog. That is 18.7 points, by our arithmetic, across 16 paired instances.
Why It Matters
This is the second follow-up on the Perplexity launches we covered this week, after Perplexity Launches 3 APIs as Its API Business Grows 4.3x and our price comparison of its search APIs.
Aravind Srinivas called it “new research on self-improving skills in Perplexity Computer” (Aravind Srinivas on X).
The Structural Read
As we read the blog, the gain comes from changing the instructions an agent loads, not the model itself, which Perplexity says plainly: personalization changes context and instructions, not model weights.
As we read it, the cap of 10 automatic skills per user shows the constraint is context space: more skills only help if the agent can still find the right one.
Perplexity research blog
“Personalization changes context and instructions, not model weights.”
Three Implications
FOR AGENT BUILDERS Perplexity reports tool call failures down 11% with automatic skills and no measurable change in cost or response time.
FOR PERPLEXITY USERS The feature is on by default for eligible users and can be turned off in settings, per Perplexity.
FOR RIVALS Perplexity says it will extend improvement to user-authored, project and organization skills next.
The Business Engineer Lens
This story maps onto the Business Engineer framework AI and the Harness Theory.
The analysis puts it this way: “It is an operating system that wraps your AI — a harness that compresses your thinking, your team’s feedback, and the model’s raw capability into something replicable.”
As we read Perplexity’s results, automatically written skills are one way that feedback gets compressed into something replicable. The analysis predates this launch and does not discuss Perplexity.
What Is Not Established
The 72.9% and 6.5% measure different things: skills Brain created for the test group against skills users added by hand in the control group. All results are Perplexity’s own and have not been independently verified.
Availability is described two ways. Perplexity’s post on X says all US consumer and enterprise subscribers; the blog says all US Max subscribers. Perplexity has not said how many users that covers.
The Bottom Line
Perplexity’s own randomized test put automatic skill creation at 72.9% of users against 6.5% creating skills by hand, about 11 times the rate by our arithmetic, with tool failures down 11% and no measured change in cost.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Every result here comes from Perplexity’s own research blog and its posts on X. The 11 times, about four skills per user and 18.7 points are our arithmetic from Perplexity’s figures.
Sources: Perplexity research blog ‘Self-improving skills’, dated 18 Nov 2026, read 19 Nov 2026 · Aravind Srinivas on X, 19 Nov 2026 · Perplexity on X, 19 Nov 2026









