Skip to content
AI Primer
release

Perplexity Computer adds automatic skill improvement for eligible US subscribers

Perplexity shipped automatic personal skill improvement, enabled by default for eligible US subscribers. An eight-week randomized test found 11% fewer tool failures.

4 min read
Perplexity Computer adds automatic skill improvement for eligible US subscribers
Perplexity Computer adds automatic skill improvement for eligible US subscribers

TL;DR

  • Brain now creates and revises personal workflow skills from past sessions, with automatic improvement enabled by default for eligible US subscribers, according to Perplexity’s announcement.
  • Tool-call failures fell 11% in an eight-week randomized production test, or about 7% excluding background runs, as the research announcement reports.
  • Built-in skill revisions run weekly, with evaluation scenarios hidden from the editing agent, a safeguard highlighted by one early analysis.

The engineering write-up includes commands for recovering session history from before context compaction. It also imposes a 10-skill cap on automatically created personal skills because the skill index itself consumes context.

The four-stage skill loop

Brain already maintains a knowledge wiki about a user’s work. Personal skills add procedural memory: reusable instructions for recurring workflows, loaded into the model’s context with model weights held fixed.

The improvement loop has four stages:

  1. Evidence: Find sessions where a skill was used, particularly failures. Stored trajectories contain messages, tool calls, reasoning and results.
  2. Reflection: Read the evidence and the full existing skill, then propose a change or abstain when the evidence is insufficient.
  3. Verification: Check the proposal using rules specific to personal or built-in skills.
  4. Activation: Apply the verified version for future sessions, which generate evidence for subsequent revisions.

Personal skill checks

Personal-skill generation runs daily in a sandbox preloaded with recent sessions, skill-usage data and the existing catalog. The agent can retrieve more session history but has no internet or connector access.

  • Creation threshold: Two independent occurrences of the same reusable behavior are required for a new skill.
  • Update threshold: One grounded occurrence can justify revising an existing skill, with at most one change per skill per run.
  • Proposal checks: The agent submits a proposal rather than editing directly. Checks cover structure, the pre-edit baseline and automatic-update eligibility.
  • Quality audit: An LLM judge checks sampled proposals for supporting evidence and useful instructions, as described in Perplexity’s thread.

At capacity, adding a skill requires retiring another and justifying why the replacement is more useful.

Built-in replay tests

A welcome bit of eval hygiene: the agent revising built-in skills receives root-cause descriptions and sanitized successful sessions, while failure scenarios stay hidden.

Weekly revision jobs target failures recurring across multiple sessions. Evidence comes from sessions eligible for training, with credentials, other secrets, attachments and media stripped.

The release gate has four steps:

  1. Generate scenarios: Reconstruct failures as self-contained tests. Connector-dependent inputs become synthetic files, allowing tests to run without connected apps or sign-in.
  2. Replay both versions: Run the original and candidate skills against the same scenarios, including controls derived from successful sessions.
  3. Judge the results: Every scenario must pass with the revised skill, and at least one must fail with the original. An LLM judge checks answers and trajectories against scenario rubrics.
  4. Approve the PR: Package the changes and evaluation results for engineer review. Approval is required before the shared skill changes.

Eight-week randomized test

Perplexity compared production users with automatic personal-skill improvement enabled against a control group with it disabled.

  • New personal skills: 72.9% of enabled users received at least one, versus 6.5% of controls who added one manually.
  • Loading in user-started conversations: 53% of enabled users had a personal skill loaded, versus 13% of controls.
  • Tool-call failure rate: 11% lower across all Computer activity; about 7% lower after excluding the system’s background runs.
  • Dissatisfaction: Medium- and high-dissatisfaction threads per user fell 7.6%.
  • Cost and response time: No measurable change.

US availability and controls

Perplexity’s November 19 rollout announcement says personal skill improvement is available to all consumer and enterprise subscribers in the US. Its November 18 engineering post specifies US Max subscribers.

Automatic improvement is on by default and can be disabled in settings. Plugins identifies Brain-authored skills and exposes individual enable/disable controls; the skills guide places installed skills under the Installed tab.

The engineering post lists user-authored, project-shared and organization skills as future expansion targets.

Share on X