Perplexity has switched on automatic skill writing for Computer, its agent product, for US Max subscribers. The system, called Brain, reads a user’s past sessions and turns repeated procedures into saved instructions the agent loads the next time a similar task appears. The change leaves the underlying model untouched; only the instructions and context around it shift, according to Perplexity’s engineering blog.

The mechanism is a four-step loop. Brain first collects earlier sessions that loaded a given skill, favoring the ones that went badly. A background agent studies the full transcripts and proposes an edit, or declines to. Automated checks, plus an LLM judge on a sample, vet the proposal before the new version goes live. Perplexity says a new skill requires two separate occurrences of the same behavior, while a single well-grounded failure can justify editing an existing one.

Two design choices stand out. Each user is capped at 10 automatically created skills, so a full library forces the agent to retire one before adding another. That limit exists because Computer loads an index of all skills at the start of every session, and a bloated index crowds the context window. The second choice is restraint: a single edit to any skill per daily run, and the agent cannot touch a skill directly. It only files proposals.

Perplexity treats its shared, built-in skills far more cautiously, since a bad edit there reaches every user. Those changes start from weekly reviews of failures that recur across sessions. Sanitized evidence is stripped of credentials and attachments, then rebuilt into synthetic test scenarios. A revision passes only if every scenario succeeds with the new version and at least one failed with the old. Control scenarios must still pass. Then an engineer approves a pull request by hand.

The adoption numbers are the strongest part of the post. Perplexity reports that Brain wrote more than 80,000 personal skills for over 20,000 users between August 1 and November 14. Over eight weeks, in a randomized test on live traffic, 72.9 percent of users with the feature enabled received at least one new skill. Among controls, 6.5 percent added one manually. That gap is the real finding: skills are a good idea that most people will not author themselves.

The performance claims are thinner. Perplexity says tool-call failures fell 11 percent across all Computer activity for the treatment group and user dissatisfaction fell 7.6 percent, while cost and response time did not move noticeably. Those figures come from the company’s own experiment, and the post names no independent reviewer. The offline benchmark behind the headline improvement, SkillLearnBench, covered nine task families and 16 paired evaluation instances, where task success rose from 43.8 percent to 62.5 percent. Sixteen instances is a small sample for a 19-point swing. The one built-in case study, an ecommerce skill, showed a 6 percent quality gain at 32 percent lower cost, judged by an automated grader.

Also unanswered is what happens when an automatic edit encodes a bad habit. A skill built from a user’s own sessions inherits that user’s shortcuts and mistakes, and the post describes no mechanism that checks personal skills against ground truth. Users can see which skills Brain authored in the Plugins tab and turn them off, which is the only stated safeguard on that side.

Perplexity plans to extend the loop to user-written skills, project-shared skills, and organization skills. Teams that already maintain hand-written agent instructions should treat the 6.5 percent control figure as a benchmark for their own habits: audit how many of those files were updated after the first week, and decide whether generated revisions plus human review would beat that record.

Reported by Perplexity (perplexity.ai blog) in November 2026.