August 2026 Summaries

Perplexity Portable Computer is presented as a local-first AI agent that runs models, conversations, tools, and task trajectories on a user’s device, using web search, external connectors, or cloud-based advisor models only with user approval. Its design pairs a compact, on-demand skill-based harness with local models such as Qwen 3.8 27B, emphasizing context efficiency, command-line connectors, self-verification, and mandatory sandboxed tool execution to reduce privacy risks and API costs. In evaluations on an NVIDIA DGX Spark, the Computer harness outperformed the general-purpose Hermes and Pi harnesses on BrowseComp web research, ParseBench-100 multimodal document understanding, and an internal 53-task knowledge-work benchmark, while generally using fewer tokens and, in two benchmarks, less time. A post-trained PPLX 27B model further improved its internal benchmark score from 82.6% to 85.4%. For more difficult coding tasks, optional escalation to a Claude Opus 5 advisor increased performance from 59.6% to 73.0% at a lower estimated cost than using the frontier model alone, though it did not eliminate the performance gap. The work argues that co-designing local hardware, models, and orchestration can make private, low-cost on-device agents increasingly viable for research, documents, data analysis, software work, and other knowledge tasks.

Perplexity has expanded its Computer agent platform with connectors to more than 20 licensed financial data sources, including Dun & Bradstreet, Guidepoint, and IBISWorld, allowing finance teams to query their existing subscribed and internal data through natural-language requests without relying on separate APIs or logins. The platform emphasizes source traceability by linking figures and answers to their underlying records, with the aim of improving research accuracy and reducing dependence on open-web information. Computer can also operate through email, enabling users to submit materials such as CIMs or data-room threads and receive outputs including models, comp sets, and portfolio updates with licensed data incorporated. The new integrations span areas such as public and private markets, credit, portfolio monitoring, regulatory filings, market news, cryptocurrency, and alternative data, supporting use cases from thesis validation and company mapping to distressed-credit analysis. Customers authenticate their own eligible partner licenses, and the connectors are available to Perplexity Pro, Max, and Enterprise users without requiring a separate data contract with Perplexity.

Perplexity has launched Portable Computer, a local version of its AI agent platform developed with NVIDIA that runs entirely on NVIDIA DGX Spark systems and is planned for future NVIDIA RTX GPU PCs. Designed for private codebases, confidential documents, and other sensitive work, it processes files, searches local data, executes workflows, and maintains planning, scheduling, routing, and indexing functions on the user’s device without consuming cloud credits. It uses Qwen 3.8 27B or Perplexity’s post-trained PPLX 27B model, with NVIDIA Nemotron 3.5 Lightning planned as an additional option, and can request permission to escalate specific tasks to cloud services for web research, connected applications, or advanced reasoning while retaining sensitive local information. The platform supports connectors such as Google Drive, Gmail, Slack, and GitHub, offers locally processed dictation through NVIDIA’s speech-recognition model, and uses sandboxed execution environments for security. Portable Computer is initially available to Perplexity Pro and Max subscribers on Linux-based NVIDIA DGX Spark hardware, with Windows support expected later.

Brain is a memory architecture for Computer agents designed to retain and apply user context across long-running work spanning many sessions. It organizes persistent information in a filesystem with a synthesized, Git-backed Markdown knowledge wiki called Brain, topical notes, and raw session histories; wikilinks connect related concepts while citations trace claims to supporting evidence. Foreground agents explore this memory through standard file operations and can use a Memory Agent to semantically retrieve and locally materialize additional files, avoiding the latency of remote filesystem access and the context overload of loading all information at once. Background “Dream” agents periodically summarize new sessions, update knowledge pages, verify evidence, and stage changes for deterministic and semantic validation before synchronization. Internal offline and production-derived evaluations reported that Brain improved correctness, evidence recall, and currentness, particularly for preferences, temporal reasoning, and information distributed across sessions, while also reducing token use, costs, and response time.

Small businesses can gain value from AI by applying it to repetitive, measurable, low-risk workflows in customer support, marketing, administration, bookkeeping, sales follow-up, and onboarding, while retaining human oversight for sensitive, strategic, legal, financial, or relationship-driven decisions. Suggested uses include drafting and organizing content, summarizing records, classifying requests, identifying anomalies, researching leads, and preparing first drafts, with clear limits against unsupervised refunds, tax advice, hiring decisions, negotiations, or public-facing commitments. Rather than selecting tools first, owners should identify a workflow bottleneck, establish a baseline for time, cost, and errors, verify data access and quality, choose software that integrates with existing systems, assign a reviewer, and run a limited pilot. AI is most suitable when review takes less time than completing the task manually, errors are recoverable, and results can be measured; return on investment should account for setup, integration, training, subscriptions, review, and corrections. Hiring remains preferable for work requiring expertise, accountability, nuanced judgment, or sustained relationship-building, though employees supported by AI may increase capacity and improve processes.

Perplexity has introduced Computer in Email, allowing users to assign tasks to its agent platform by emailing, forwarding, or copying [email protected] into a conversation, with results returned in the same thread. The system verifies senders, processes email threads and attachments using each user’s existing permissions and connected tools, and can return deliverables such as Excel models, PDFs, and presentations. Each request creates a full Computer session accessible on web and mobile, including its usual audit trail, while drawing on persistent memory of team terminology, preferred formats, prior assumptions, and earlier decisions. The feature is intended for email-based workflows such as private-equity modeling, legal redline reviews, and investment-banking valuation analysis, extending Computer’s availability beyond prior integrations with Slack and Microsoft Teams. It is available immediately to all Computer users, although reply-all functionality for Enterprise customers is planned for a future release.

Perplexity is transitioning Sonar customers to its Agent API, a unified programmable endpoint for web search, URL retrieval, code execution, MCP connections, and finance and people search, with tools composed dynamically through its Search as Code architecture. The API replaces Sonar’s fixed model tiers with configurable presets ranging from fast to wide-research, each combining a model, system prompt, tool settings, reasoning effort, and token budgets that can be updated as frontier models improve while allowing parameter overrides. Perplexity says benchmark testing shows the presets deliver higher quality at lower costs, including an approximately sevenfold BrowseComp improvement for the low preset over Sonar Pro at about $0.05 per query, while potentially matching Sonar Deep Research more cheaply. Existing Sonar models map to corresponding presets, with xhigh adding support for more open-ended agentic tasks, and users are advised to migrate because Sonar endpoints will retire on September 27, 2026, though customers with contractual commitments retain access through their current terms.

Perplexity has introduced Computer for Builders, a set of agent-driven tools aimed at solo founders and small engineering teams seeking to manage software development and business operations with less manual work. Computer uses an orchestrator spanning more than 15 AI models and integrates with services including GitHub, Datadog, Stripe, Supabase, Slack, Vercel, and hundreds of connected workplace tools to write and review code, deploy applications, monitor production, query data, track payments and fraud, and deliver recurring reports. The platform is positioned to support end-to-end workflows such as diagnosing and fixing bugs, implementing subscriptions, adding AI features, and identifying churn patterns before measuring the results of retention efforts. Computer for Builders is available to Perplexity Pro and Max subscribers using Computer, with a free trial option for new users.