Perplexity has extended Brain, the personalization system inside its Computer agent, so it can write and update agent skills on its own by learning from a user’s past sessions. Between 1 August and 14 November 2026, Brain wrote more than 80,000 personal skills for over 20,000 users, Perplexity said.
Quick Take
- What’s new: Brain now creates and refines reusable “skills”, step-by-step instructions for recurring tasks, from evidence in earlier sessions.
- Scale: over 80,000 personal skills for more than 20,000 users, plus more than 100,000 automatic updates to over 30,000 existing skills since 4 September.
- Test result: in an eight-week randomized test, tool-call failures fell 11% and dissatisfied threads per user fell 7.6%, with no measurable change in cost or response time.
- Availability: personal skill improvement is on by default for US Max subscribers and can be turned off.
- Guardrails: changes to built-in skills shared by all users need replay tests and an engineer’s approval.
In plain English
If you ask an AI agent to do the same kind of job every week, say a report in a certain format, you usually have to explain your preferences again and again. Perplexity’s agent now notices those patterns and writes itself a short instruction sheet for that task. Next time, it loads the sheet automatically. If a sheet keeps falling short, the agent rewrites it.
What happened
In a post dated 18 November 2026 and shared by CEO Aravind Srinivas, Perplexity explained that “we’re extending Brain to automatically create and update agent skills.” Brain already kept a knowledge wiki about each user and their work. Now it also turns repeated procedures into skills, and Perplexity applies a similar loop to the built-in skills every user gets, based on aggregate patterns.
Computer’s personalized self-improvement agent, Brain, now automatically creates self-improving agent skills for you.
Perplexity (@perplexity_ai) · 19 November 2026 · View on X
Brain learns how you work from past sessions and memory, then writes and updates personalized skills for future tasks.
In the eight-week randomized test on production traffic, 72.9% of users with the feature got at least one new personal skill, while only 6.5% of the control group added one by hand. A personal skill was loaded in 53% of conversations started by the test group, against 13% for the control group.
Background
Perplexity Computer is the company’s general-purpose agent product, which recently added a multiagent mode where up to 16 agents check each other. Skills are a common way to steer agents without retraining a model: they change the agent’s instructions and context, not the model’s weights.
How it works
- Four-stage loop: evidence (find sessions where a skill fell short), reflection (an improvement agent proposes changes), verification (checks that the proposal is sound) and activation (the new version is used from then on).
- Evidence rules: according to Unite.AI’s summary of the post, a new skill needs two independent occurrences of the same behavior, while one grounded occurrence can justify an update. The improvement agent has no internet or connector access.
- Caps: automatically created skills are limited to 10 per user; at that limit, the agent must retire one and justify the swap.
- Built-in skills: a weekly job uses sanitized evidence, turns recurring failures into test scenarios and requires every scenario to pass with the new skill in paired replay before an engineer approves the change.
- Transparency: the Plugins tab marks skills written by Brain and lets users switch them on or off.
Why it matters
For users, agents that learn your routines could cut the time spent re-explaining preferences, which is one of the main frustrations with today’s assistants.
For developers building agents, Perplexity’s design is a concrete example of “learning from experience” without fine-tuning: small, auditable text changes, gated by tests and human review where many users are affected.
For companies, the open question is control. A skill that silently changes how an agent handles a workflow needs to be visible and reversible, which Perplexity addresses with the Plugins tab and the opt-out.
What’s still open
- All figures come from Perplexity’s own experiments.
- Perplexity plans to extend the system to user-written skills, project skills and organization skills, where automatic changes affect whole teams.
- When the feature reaches users outside the US and other plans.
Sources
- Perplexity, “Self-improving skills,” Perplexity Engineering, 18 November 2026
- Perplexity and Aravind Srinivas on X, 19 November 2026
- Unite.AI, report on Perplexity Brain, 19 November 2026
i10x Perspective
Memory told agents who you are; self-written skills tell them how you work. That is a meaningful step toward agents that improve with use, and Perplexity’s caps and review gates are the right instinct. More on agent products in Agents.






