July 2026 Summaries

Computer has introduced Projects, replacing Spaces as collaborative hubs for long-running work that combine persistent shared files, task history, instructions, automations, and optional self-improving memory through Computer Brain. Projects allow users and agents to work on the same evolving documents, spreadsheets, folders, and connected sources, while Brain reviews prior sessions and files to preserve context, update project summaries, and improve future tasks. Users can upload or import files, connect more than 400 services including Google Drive, Notion, Linear, Snowflake, and GitHub, and link Slack or Microsoft Teams channels so conversations and tasks contribute to project context. Project-level instructions define overarching rules and defaults, while reusable Skills provide standardized workflows for recurring work. Collaboration is designed to maintain privacy, as shared project resources do not grant access to teammates’ personal memory or connected accounts, and connector requests use each individual’s credentials and permissions. Existing Spaces migrate automatically to Projects, which are available to all Computer users.

Perplexity has open-sourced Numbat, a security suite for monitoring and controlling AI coding agents on enterprise endpoints, arguing that agent risks extend beyond prompt injection to “accidental meltdowns,” where autonomous agents pursue legitimate goals through harmful, unauthorized actions after encountering ordinary errors or constraints. Numbat integrates with common agent harnesses through lifecycle hooks, filesystem-based session artifacts, and OpenTelemetry telemetry, enabling real-time policy enforcement, action blocking, normalized forensic timelines, and local-first data collection. Its built-in and customizable rules detect behaviors and sequences associated with secret access, data exfiltration, privilege escalation, lateral movement, and destructive system changes, while supporting retrospective investigation of past sessions. Perplexity deploys Numbat across its own fleet for tools including Claude Code, Codex, OpenCode, and Pi, combining centralized telemetry review with automated rule-improvement workflows subject to human approval. Available as an open-source project for macOS, Linux, and Windows, Numbat is positioned as a system-level safeguard for securing autonomous agents across varied local, cloud, and hybrid environments.

Perplexity has launched Personal Computer for Windows, extending its AI agent platform to Windows 10 and 11 devices for Pro, Max, and Enterprise subscribers. The service is designed to work across local files, Microsoft 365 applications, connected business tools, and the web, enabling users to create and edit Office documents, organize files, research information, update presentations and spreadsheets, and complete multi-step workflows from a single interface. It builds on prior integrations with Microsoft 365 and Teams, while adding access to more than 400 external tools through App Connectors, including platforms such as Snowflake, Salesforce, and HubSpot. Users can initiate tasks from a phone and continue them on a desktop, use voice commands, and allow the agent to handle workflows involving applications, email, web research, and file management. Perplexity says the platform provides cited information, requires user alerts for sensitive actions such as sending emails or deleting files, creates files in a secure sandbox, and maintains auditable actions without training on company data.

Perplexity has introduced Model Council in its Computer product, allowing users to assemble panels of two to eight AI models from providers including OpenAI, Gemini, Anthropic, and open-source developers to evaluate complex questions. Users can choose the models and analysis depth, while the feature synthesizes consensus, disagreements, unique insights, confidence levels, and underlying assumptions before using Computer to create practical outputs such as reports, presentations, PDFs, memos, spreadsheets, and documents. Designed for ambiguous, high-stakes decisions rather than simple summaries, Model Council supports applications in legal analysis, finance, corporate strategy, business planning, engineering, and personal decisions, especially when users provide specific constraints, tradeoffs, timelines, and personas. It can be accessed through Computer’s model-selection menu or by directly requesting a Model Council query, and is available on web and mobile for Perplexity Max and Pro users.

Perplexity has introduced SPACE, a sandbox platform for its Computer agent system that aims to combine strong security, efficiency, and functionality for AI agents performing long-running tasks involving code, files, networks, and sensitive information. SPACE creates ephemeral Firecracker microVM sandboxes that are destroyed after tasks finish, while session-level rolling snapshots preserve memory and files so work can be paused, resumed, branched, or restored for up to a week. Its architecture separates responsibility among a stateless Control Plane, node-level services that manage credentials and network access outside the sandbox, and isolated in-sandbox processes, limiting a compromised agent’s access to only its authorized task. The platform also supports credential isolation, controlled protocol forwarding, tenant separation, encrypted storage, customer-managed encryption keys, multiple infrastructure backends, and on-premises or offline deployment. Built in 10 weeks and used internally for two months, SPACE is now available to all Computer users and has reportedly handled millions of sandbox creations, with early NVIDIA Vera CPU tests indicating faster agent workflow and concurrent sandbox startup performance.

Perplexity introduces SPACE, a sandbox platform designed for long-running, stateful AI agent workflows that require secure access to code execution, filesystems, networks, and multi-step task environments. Unlike conventional container-based systems, SPACE uses per-sandbox virtual machines to provide separate guest kernels and stronger isolation, while a controlled in-VM daemon manages filesystem, process, and platform interactions. Its three-layer architecture combines a stateless control plane, node-local services for lifecycle, storage, credentials, and networking, and isolated VM-based sandboxes. Security measures include restricted communication paths, centralized egress enforcement, external credential storage and injection, audit controls, and customer-managed encryption keys. SPACE supports agent features such as rapid creation, persistence, pausing, recovery, rollback, forking, and migration through disk and full-VM snapshots, while btrfs copy-on-write storage, warm template pools, and deduplicated image downloads reduce resource use and startup time. Perplexity reports that SPACE, now powering all Computer sessions, handled millions of sandbox creations and improved median sandbox creation latency from 185 milliseconds to 60 milliseconds, with the platform intended to expand to additional environments including Windows and local machines.

WANDR (Wide ANd Deep Research) is an open benchmark and evaluation harness containing 500 realistic knowledge-work tasks that test whether research agents can both discover large sets of qualifying entities and gather evidence for every required claim, complementing the DRACO benchmark’s focus on long-form reports. Its tasks model nested research structures such as finding companies, employees, and supporting URLs, with a median requirement of 50 members and 245 source-backed records per task, totaling more than 170,000 records across the release. Rather than relying on fixed answer keys, WANDR grades each submitted record by re-fetching cited sources and evaluating page usability, claim scope, excerpt fidelity, and evidentiary support, then reports soft and hard precision, recall, and F1 scores that distinguish partial from fully completed research paths. In tests of six production systems, Perplexity’s Search as Code achieved the highest main-run results at 0.363 soft F1 and 0.133 hard F1, indicating that comprehensive, evidence-backed research remains difficult even for leading systems. Results showed that systems commonly provide some valid findings but struggle to meet requested coverage, complete deeper hierarchies, and supply excerpts that support all aspects of their claims; discovery shortfalls and evidence construction were particularly significant bottlenecks. The public tasks, verifier, diagnostics, and technical report are intended to support development, evaluation, and potentially reinforcement-learning training for agents that can better plan, verify, and complete broad research assignments.