Perplexity: Brain wrote more than 80,000 skills from over 20,000 users’ past tasks
Perplexity’s Brain builds skills, ready-made work instructions, for users of the Computer agent from their past sessions. In the company’s own experiment, tool call failures fell by 11 per cent. That figure, however, also includes the system’s own background runs.
Perplexity announced on 18 November that Brain now writes and updates skills itself. Brain is the personalisation system for the company’s Computer agent. A skill is a reusable instruction that the agent loads for a recurring task, and it can include scripts. The model weights do not change; instead, learning takes the form of instructions the agent can use.
Brain wrote more than 80,000 skills between 1 August and 14 November and has made more than 100,000 updates since 4 September, the company’s blog said. OpenTools repeats the same figures from Perplexity’s blog via Unite.AI, so it does not confirm them independently.
Once a day, an improvement agent reads the user’s recent sessions in a sandbox with no internet access and no access to connected services. A new skill is created only when the same way of working appears twice independently. One observation is enough for an update. Automatic checks ensure that a proposal can be deployed. However, an evaluator powered by a language model reviews only a sample of the proposals. Automatically created skills are capped at ten, because their list is loaded at the start of every session.
In an eight-week randomised experiment, Perplexity compared a test group using the feature with a control group that had it switched off. In the test group, 72.9 per cent received at least one new skill, while 6.5 per cent of the control group added one themselves. Tool call failures fell by 11 per cent. That figure includes the system’s own background runs; without them, the reduction was about 7 per cent. The number of conversations per user showing medium or high dissatisfaction fell by 7.6 per cent. Cost and latency showed no change. The company said the results are preliminary.
On the SkillLearnBench benchmark, the share of successful tasks rose from 43.8 per cent to 62.5 per cent. The benchmark had only 16 test cases, however, each run both with and without the skill.
Changes to skills shared by all users take effect only after the old and new versions have been run on the same failure cases and an engineer has approved the pull request.
The feature is on by default for US Max subscribers, but can be turned off. Max costs $200 a month, OpenTools reported. The Plugins tab shows the skills Brain has written and lets users choose which of them are enabled.
Read more: How an agent works · Agent map