Blogs
Perplexity logo

Perplexity

Engineering blog · perplexity.ai/hub/blog

4posts0this week0watchers

Feature trails

3
Feature trailActive

Agentic Memory System

Perplexity has developed and open-sourced WANDR, a benchmark and evaluation harness for research agents that must search wide and deep. WANDR comprises 500 realistic, challenging data-collection tasks, designed to evaluate an agent's ability to discover a large set of qualifying entities and investigate each sufficiently to support requested claims with evidence. The benchmark uses a flexible, composable qualification key hierarchy and a reference-free, evidence-verified grading system. Initial evaluations show that while partial progress is common, complete coverage remains a significant challenge for current systems, with Perplexity's Search as Code leading in performance.

3 posts · 2026

Timeline

Feature trailActive

Computer in Email Integration

Computer now integrates directly into email clients, allowing users to initiate tasks, process thread content and attachments, and receive outputs within email conversations. This extends the 'Computer' agent platform's reach into a core knowledge work surface, enabling tasks like model building and issue summarization directly from the inbox. The integration leverages existing Computer sessions, memory, and user permissions, with outputs delivered as attachments or within the thread. Future enhancements include 'reply all' support for Enterprise.

1 post · 2026

Timeline

Feature trailActive

Agent Security Suite

Perplexity has developed and open-sourced Numbat, an agent security suite designed to prevent, detect, and mitigate security incidents arising from AI agents, particularly on client endpoints. Numbat integrates with common agent harnesses to enforce security rules, monitor agent behavior, and provide forensic capabilities. It addresses novel security challenges beyond traditional prompt injection, focusing on 'accidental meltdowns' where agents pursue goals in unintended and harmful ways due to environmental errors or complex task execution. Numbat offers local-first detection and agent-agnostic forensics through hooks, session artifact analysis, and OTLP telemetry.

1 post · 2026

Timeline