How the loop works
Perplexity has extended Brain, the personalization system inside its Computer agent, to write and update reusable agent skills on its own. The system learns from a user’s session history and memory, turning repeated workflows into skills the agent can reuse.
Brain improves a skill through a four-stage loop: evidence, reflection, verification and activation. It finds past sessions where a skill fell short, an improvement agent proposes an update, checks confirm the change is well formed, and the new version goes live for future sessions.
Personal skills come from one user’s own sessions, with a creation needing two independent examples of the same reusable behavior. Built-in skills, shared by every user, face a higher bar: paired replay tests and a human engineer’s approval before any change ships.
The system caps personal skills at ten per user, retiring weaker ones when a better skill earns the slot. Personalization changes context and instructions, Perplexity says, not the model’s weights — the agent gets better at a user’s work without being retrained.
What the tests showed
Between August 1 and November 14, Brain wrote more than 80,000 personal skills for over 20,000 users and applied more than 100,000 updates to 30,000 existing skills. The numbers come from Perplexity’s own production data, and no independent check of them exists yet.
In an eight-week randomized test, 72.9 percent of users with the feature on gained at least one new skill, against 6.5 percent of controls who added one by hand. Tool-call failures fell 11 percent, and dissatisfied threads per user fell 7.6 percent, with no change in cost or speed.
Offline tests point the same way. In a SkillLearnBench experiment, task success rose from 43.8 percent without generated skills to 62.5 percent with them, across sixteen paired cases. One built-in ecommerce skill edit lifted quality 6 percent while cutting cost by nearly a third.
The work sits inside Computer, Perplexity’s agent product, which the company has been expanding quickly — it shipped three new APIs for the agent business this week. Brain is the memory layer underneath it, deciding what the agent remembers and reuses.
Our Read
The interesting claim here is not the skill count but the loop. Most agents forget everything between sessions; Brain is Perplexity’s attempt to make the agent compound its experience instead, with checks and human approval keeping the self-editing from drifting off course.
The test numbers are strong, but they are Perplexity’s own, from an experiment it designed and ran. The 72.9 percent figure measures skill creation, not task success — and the real test is whether users notice their agent getting better, not whether its library grows.
Why It Matters
If agents are going to do real work, they have to learn how a specific user works — and keep learning without being rebuilt each time. Brain is Perplexity’s answer to that problem, and its measured results make it one of the more concrete self-improvement claims in the agent race.
Background
The feature is live now for all United States Max subscribers, switched on by default. It lands as Perplexity pushes Computer deeper into multi-agent work, with the company testing teams of rival AIs against each other and measuring the results.
Source: Perplexity Engineering Blog


0 Comments