AI News
AI News AgentProductPerplexity4 min read

Perplexity Automates AI Skills in Computer

Perplexity has introduced a system that allows Computer to create and update skills based on each user’s previous tasks. Early tests reduced tool-call failures by 11% and dissatisfaction by 7.6%, although the feature is currently limited to Max subscribers in the United States.

Perplexity is making its Computer agent learn from the tasks it performs so it can repeat each user’s workflows more effectively. The system analyzes past sessions, detects patterns, and creates or modifies reusable instructions called skills.

The idea is simple: if you always correct the agent in the same way, you will not have to explain the process again in every conversation.

How the Agent Learns

Computer is part of Perplexity’s platform and uses Brain, its personalized memory system. Brain maintains a structured wiki about the user and their work, connected to evidence from their sessions. Agents running in the background update that information over time.

Skills add a practical layer to that memory. They are instructions that guide the agent through specific tasks, such as reconciling payroll, checking software, or preparing a report. They can also include scripts for repetitive processes that are better expressed through code.

The improvement system follows four steps:

  • Evidence: searches for sessions in which a skill was used, especially when there were errors, frustration, or unnecessary work.
  • Reflection: an agent analyzes the full conversations and proposes a change.
  • Verification: automated checks and reviews with another model assess whether the proposal is justified.
  • Activation: if it passes the checks, the new version becomes available for future tasks.

Perplexity does not only look at whether the final result was good or bad. Computer stores conversations, tool calls, and results in a format that makes it possible to reconstruct each session. This allows the agent to identify the exact point where an instruction failed.

Personal and General Skills

Personal skills are created from a user’s history. To create a new skill, the system requires the same reusable behavior to have appeared in two independent instances. Updating an existing skill requires just one well-supported occurrence.

The user does not simply receive an applied change: the system generates a proposal containing the modifications, the evidence, and the reason for the adjustment. It also limits the system to 10 automatically created skills per user. If the library reaches its limit, the user must remove one before adding another, preventing too many instructions from saturating the agent’s context.

Perplexity also improves its built-in skills, which are available to all users. In this case, it uses sanitized sessions to remove credentials, files, and other sensitive data. It then turns recurring failures into synthetic test scenarios.

A revision can move forward only if the new version passes all test scenarios, improves at least one scenario that failed with the previous version, and does not break cases that were already working. Finally, an engineer must approve the change before it reaches the official version.

Results So Far

Between August 1 and November 14, Brain created more than 80,000 personal skills for more than 20,000 users. It also applied more than 100,000 updates to more than 30,000 existing skills between September 4 and November 14.

In an eight-week randomized trial:

  • 72.9% of users with automatic improvement received at least one new skill, compared with 6.5% of the control group, which had to add them manually.
  • The failure rate for tool calls fell 11% across all Computer activity and by nearly 7% when excluding the system’s internal tasks.
  • Threads with medium or high user dissatisfaction decreased by 7.6%.
  • No measurable change was observed in cost or response time.

In a smaller trial, the share of tasks completed with skills generated from previous sessions rose from 43.8% to 62.5%. It is a useful signal, although it comes from only 16 paired evaluations.

What This Means for You

For now, automatic skill improvement is available to Max subscribers in the United States. It is enabled by default for eligible users, but can be turned off. The Plugins tab shows the available skills, indicates which ones were created by Brain, and lets you choose which ones to use.

This does not mean that Computer learns anything without oversight. The system needs repeated patterns, maintains a small library, and subjects general skills to testing and human review. The next step will be extending the mechanism to user-created skills, shared projects, and organizations.

The direction is clear: agents will stop relying only on a general model and start accumulating procedural memory. The important question will be whether that memory remains useful and accurate, because an instruction learned incorrectly can also be repeated automatically.