Skip to main content
Back

Perplexity Brain Now Builds and Refines Its Own AI Skills Automatically

Perplexity's Brain now mines your past sessions to auto-write reusable agent skills, cutting tool-call failures by 11% in production tests.

  • Perplexity's Brain now auto-writes reusable agent skills from your past sessions and memory.
  • Four-stage loop: evidence, reflection by background agent, verification checks, activation.
  • Production test: 11% fewer tool-call failures, 7.6% less user dissatisfaction, no cost change.
  • 72.9% of enabled users got a new skill vs 6.5% who added one manually.
  • Per-user cap of 10 auto-created skills; built-in skills gated by human PR review.
  • Live now for US Max subscribers, on by default, toggleable in Plugins settings.

Perplexity has expanded its Brain announcement into a continual-learning system for Computer, its agent. Brain reviews prior sessions, identifies repeated workflows, and writes or refines reusable instruction sets called skills. Those skills give future sessions procedural memory about a user’s preferred tools, resources, and methods.

A skill can contain prose instructions, references, and executable scripts for procedures that code can handle reliably. This saves Computer from reconstructing the same process during every session. Users previously had to create and maintain skills manually; Brain now automates both tasks from session evidence.

A loop that learns from past runs

Each personal or built-in skill moves through a four-stage improvement loop:

  1. Evidence: Brain retrieves sessions in which the skill was used, prioritizing runs where the result fell short. An improvement agent examines that evidence and identifies a possible revision.
  2. Reflection: A background agent runs in an isolated sandbox, reads the full session trajectory, and drafts a specific change to the skill.
  3. Verification: Deterministic checks validate the proposal’s structure. Built-in skills also undergo paired replay, which compares the existing and revised versions on held-out failure scenarios.
  4. Activation: An approved revision becomes available to future sessions, which then produce evidence for later updates.

The reflection stage uses complete trajectories containing messages, tool calls, reasoning, and results. Perplexity stores these records in a provider-agnostic format, allowing the improvement system to process runs from different model providers through one interface. A command-line toolkit can decode compressed trajectories, show which skills were loaded, and inspect individual turns to locate a failure.

Personal skills move faster than shared ones

Perplexity applies different evidence and release requirements according to a skill’s reach:

Requirement Personal skills Built-in skills
Audience One user All users with access to the skill
Creation evidence Two independent occurrences of the same reusable behavior Failure scenarios gathered and reconstructed for testing
Update evidence One grounded occurrence can justify a revision Paired replay must show the revised version passing every scenario while the original fails at least one
Release process Daily per-user improvement cycle with automated verification Candidate revision, deterministic checks, paired replay, human review, and engineer approval
Capacity Up to 10 automatically created skills per user No limit disclosed

The 10-skill personal limit keeps stale instructions from consuming the context window. Once the library reaches capacity, the agent must retire an existing skill and explain why the replacement is more useful.

Built-in revisions remain candidates until they clear stricter release gates. A scenario-generation agent reconstructs failures synthetically, allowing replay tests to run independently of connected apps. An engineer approves the resulting pull request only after the revised skill passes every test case and the original version fails at least one.

Production tests show fewer failures

Perplexity ran an eight-week randomized experiment on live production traffic, comparing users with personal-skill improvement enabled against a control group. The company reported substantial differences in both skill adoption and usage:

  • 72.9% of users with improvement enabled received at least one new personal skill, compared with 6.5% who manually added one in the control group.
  • 53% of the treatment group loaded a personal skill in a conversation they started, compared with 13% of the control group.
  • Turns in the treatment group loaded personal skills about 3.6 times as often.

The production experiment and separate skill evaluations produced the following quality results:

Evaluation Reported result
Tool calls across Computer Failure rate fell 11% overall and about 7% when background runs were excluded
User dissatisfaction Medium- or high-dissatisfaction threads per user fell 7.6%
Cost and latency No measurable change
SkillLearnBench Task success rose from 43.8% to 62.5% across nine task families, an increase of 18.7 percentage points
E-commerce built-in skill An automated rubric measured a 6% quality increase and a 32% cost reduction after one revision

The e-commerce result used automated grading, while the production experiment relied on operational failures and user-dissatisfaction signals. These measurements come from separate evaluations and describe different aspects of performance.

Between August 1 and November 14 of the reported period, Brain created more than 80,000 personal skills for over 20,000 users and made more than 100,000 automatic updates. Software engineering represented 11% of skills, followed by research and briefings at 10.1%, sales and operations at 9.9%, and finance and investing at 9.9%. Other examples included payroll reconciliation, figure design, and pull-request triage.

Adaptation happens in the instruction layer

Brain changes the context and instructions supplied to the agent while the underlying model weights remain fixed. This approach removes the training pipeline and weight-level catastrophic-forgetting risk associated with repeated fine-tuning. It introduces a separate risk that instructions may become stale, overly specific, or contradictory, which Perplexity addresses through evidence thresholds, library limits, deterministic validation, sampled audits by an LLM judge, replay tests, and human review.

The architecture separates interactive work from background improvement through several components:

  • A shared modular harness gives interactive and improvement agents access to the same core capabilities.
  • Provider-agnostic trajectory storage records sessions in a common format across model providers.
  • Skill-load events and structured diagnostics show which instructions influenced a run and where failures occurred.
  • A separate worker fleet scales reflection and evaluation independently of user-facing response latency.

Perplexity’s earlier posts on skills architecture and memory design provide additional detail on the underlying systems.

US Max users get access first

Personal skill improvement is available to Perplexity Max subscribers in the United States. The feature is enabled by default for eligible accounts and can be disabled. The Plugins tab lists available skills, identifies those created by Brain, and provides controls for enabling or disabling individual skills. It also shows which recurring behaviors Brain has converted into reusable instructions.