Updated 12 hours ago
Perplexity Brain Creates 80,000 Agent Skills for 20,000 Users From Experience

Agent Personalization

Perplexity Brain Creates 80,000 Agent Skills for 20,000 Users From Experience

Perplexity's Brain personalization system has written over 80,000 personal skills for more than 20,000 users, learning workflows from experience. The system reduced tool‑call failures by 11% and dissatisfaction threads by 7.6% in production testing.

80,000 Personal Skills, 20,000 Users, 3.5 Months

Perplexity announced on November 18 that Brain, the personalization system for its Computer agent, can now create and update reusable agent skills automatically. Between August 1 and November 14, 2026, Brain wrote more than 80,000 personal skills for over 20,000 users — skills that capture workflow‑specific preferences and procedures a general‑purpose model cannot know in advance, Unite.AI reported, drawing from Perplexity's announcement.

Skills package reusable instructions and supporting resources, sometimes including scripts. The Computer agent's system makes skills discoverable and loads their instructions when the agent requests them. Users can also specify skills directly in prompts. Crucially, personalization changes only the agent's context and instructions — not the underlying model weights.

  • 80,000+ skills created Written automatically for 20,000+ users between August 1 and November 14, 2026
  • 100,000+ updates Automatic updates to 30,000+ existing skills between September 4 and November 14, 2026
  • 72.9% adoption Users with auto‑improvement enabled received at least one new skill vs. 6.5% of controls who manually added one
  • 63% retention Among users who received a Brain‑written skill, 63% used one again within four weeks

How Skills Are Created: Four‑Stage Improvement Engine

The skill creation and improvement process runs through four stages: gather evidence, reflect on it, verify a proposal, and activate the change. The agent logs skill loads and files structured diagnostics when it detects user frustration, broken behavior, or clear inefficiency, per Unite.AI, citing Perplexity's official blog.

Critically, stored session trajectories preserve full messages, tool calls, reasoning, and results — allowing the improvement agent to inspect exactly where a procedure went wrong rather than relying on a vague outcome or short summary. A new skill requires two independent occurrences of the same reusable behavior, while one grounded occurrence can support an update. Each proposal is limited to one change per skill per run, with deterministic checks and an LLM judge auditing sampled proposals.

Measurable Impact: Fewer Failures, Less Frustration

The production test results show real improvements. Perplexity reported an 11% reduction in tool‑call failures across all Computer activity, and about 7% after background runs were excluded. Medium- or high‑dissatisfaction threads per user fell by 7.6%, with no measurable change in cost or response time, according to Unite.AI.

In an eight‑week randomized production test, 72.9% of users with automatic improvement enabled received at least one new personal skill, versus just 6.5% of controls who manually added one. A personal skill was loaded in a user‑started conversation for 53% of the enabled group versus 13% of controls. These are substantial gaps that suggest automatic skill creation meaningfully changes how users interact with the agent.

SkillLearnBench: 43.8% to 62.5% Task Success

Perplexity also tested Brain on SkillLearnBench, a benchmark of 20 tasks across six real‑world categories with 100 verified instances. In a narrower offline experiment covering nine task families, Computer generated one skill for each, and task success across 16 paired instances rose from 43.8% without skills to 62.5% with generated skills — a 43% relative improvement, Unite.AI reported.

The company noted this experiment covered only a subset of the full benchmark suite, but the direction is clear: agent skills that capture reusable workflows substantially improve task completion rates.

Safety Guardrails on Automatic Updates

Perplexity built multiple safety layers into the skill update system. Built‑in skill updates require paired replay testing and engineer approval before any change takes effect. In daily personal‑skill runs, a background improvement agent receives recent sessions, skill‑usage data, and the existing skill catalog — but has no internet or connector access. It can retrieve more of the user's session history as needed but operates in a constrained environment.

Personal skill improvement is available to all US Max subscribers, enabled by default but can be turned off. The Plugins tab identifies Brain‑authored skills and lets users choose which skills are enabled, giving users visibility and control over what the system has learned about their workflows.

What This Means for Builders

Perplexity Brain represents a significant step toward agents that learn from experience rather than requiring explicit programming for every workflow. For developers building on or competing with agent platforms, several patterns stand out:

1. Context over weights. Brain personalization changes instructions and context, not model weights. This is a pragmatic approach — it avoids the risks and cost of fine‑tuning while still delivering personalized behavior. Any builder working on agent personalization should consider this architecture first before reaching for weight‑level changes.

2. Structured diagnostics matter. The 11% reduction in tool‑call failures came from logging structured diagnostics when the agent detects frustration or broken behavior. This isn't magic — it's good instrumentation. Builders should invest in similar diagnostic logging for their own agent systems.

3. Adoption happens fast when it's automatic. The 72.9% vs. 6.5% gap between automatic and manual skill creation is a clear signal: users won't create skills manually. The personalization has to be invisible and automatic to achieve meaningful adoption.

Share this article

PostShare

More on This Story

Related News