Perplexity Computer adds self-improving agent skills
Perplexity unveiled an automated skill-synthesis engine for Brain, the memory system powering Perplexity Computer, enabling background agents to write and refine reusable skills from user session trajectories. The system detects recurring procedural patterns and execution bottlenecks, cutting tool-call failures by 11% in production testing without manual prompt engineering.
Transforming agent memory from passive chat recall into self-improving procedural skills solves one of the biggest bottlenecks in autonomous workflows: repetitive prompt babysitting. Distilling raw execution trajectories into reusable, verified instructions provides a concrete blueprint for how agentic systems can compound capability over time.
- –Procedural over declarative memory: While typical chatbots store static user preferences, Brain analyzes tool-call traces, error states, and user corrections to codify executable workflows.
- –Evidence-based mutation: A four-stage improvement engine (gather, reflect, verify, activate) prevents prompt regression by requiring multi-session evidence before proposing or activating skill updates.
- –Measurable reliability gains: Production trials demonstrated an 11% reduction in tool failures and a 7.6% drop in dissatisfied user sessions without added latency.
- –Convergence on modular skills: Packaging learned behaviors into discoverable skill definitions demonstrates the ecosystem coalescing around composable, human-auditable agent instructions rather than opaque fine-tuning.
DISCOVERED
1h ago
2026-11-19
PUBLISHED
3h ago
2026-11-19
RELEVANCE
AUTHOR
AravSrinivas