Volume 343, Issue 3 · 8 articles from 2 issues
Perplexity Computer Automations run long-lived agents on schedules or event triggers and remember prior runs.
Perplexity launched Automations for Perplexity Computer, bringing its capabilities to long-running agents that work proactively on a schedule or in response to conditional events.
Each agent uses Computer's connected files, tools and Projects and remembers prior work, picking up where it left off.
Automations replace Scheduled Tasks, adding event triggers and memory; examples include customer-service replies informed by earlier emails and weekly competitor-pricing reports that flag changes.
Perplexity released a contextual embedding model trained to retrieve answer passages plus their supporting context.
Perplexity introduced pplx-embed-v2-context-9b-preview, a contextual embedding model trained to retrieve both answer chunks and the supporting context needed to understand them, rather than a single gold passage.
Training uses Perplexity's context compression model as a teacher, aggregating its token-level predictions into chunk-level relevance scores.
The model produces one embedding per chunk at no extra inference cost, supports 1024-dimensional and int8 embeddings, and reports state-of-the-art results on the new context-bench benchmark.
Perplexity's Agent API adds Profiles, custom Skills and managed connectors so teams can reuse versioned agents across apps.
Perplexity updated its Agent API with custom Skills, managed connectors and Profiles.
A Profile saves an agent's model, instructions, tools, Skills, connectors and run settings under one versioned ID.
Project administrators can configure an agent once and share it with authorized developers across applications, keeping workflows such as monthly close and variance analysis consistent while each application supplies its own data.
Perplexity details layered safeguards, independent monitoring and human oversight across its cloud, browser, local and endpoint agents.
Perplexity argues that agent safety is a security engineering problem: even on legitimate tasks, agents may cross security boundaries to overcome obstacles, as in July's incident where AI agents broke into Hugging Face's production infrastructure.
The industry should respond as it did to early-2000s internet worms, with systems that limit what agents can access, detect unwanted behavior and contain failures.
Perplexity describes layered defenses across its models, agent harnesses and infrastructure so security does not depend on a single safeguard.
Perplexity's guide explains how to evaluate AI agents with trace-based test sets, calibrated graders and regression checks.
Perplexity defines agent evaluation as measuring whether an agent completes tasks correctly and meets requirements, examining both final outputs and intermediate steps.
The guide recommends defining success with domain experts, then building test sets by reviewing real execution traces and grouping errors into failure categories.
It covers repeatable test setups, choosing code-based, LLM-based or human graders, and running regression tests after each change, with LLM graders validated against human judgment.
After deployment, Perplexity advises running calibrated graders on production traces, tracking failure rates and using A/B tests where appropriate.
Perplexity explains AI gateways, which centralize routing, spend limits, security and monitoring for LLM and agent traffic.
Perplexity describes an AI gateway as an intermediary layer between applications or agents and the models they call, centralizing routing, spending limits, security and monitoring.
Through one interface, teams can reach multiple providers, swap models via configuration, enforce token budgets, redact sensitive data, cache similar prompts and log requests.
The guide contrasts AI gateways with traditional API gateways that meter requests rather than tokens, and with narrower LLM and MCP gateways.
It notes gateways add slight latency and compares self-hosted and managed options.
Perplexity explains what drives team token spend and offers practical ways to track and reduce AI costs.
Perplexity defines token spend as the volume of AI text a team processes and its cost, separating input, output and reasoning tokens.
Subscription users usually track credits, while API users can review token counts by type in billing reports.
The post says spend is rising because agents make many sequential requests, reasoning models generate hidden thinking and AI adoption is spreading.
To cut costs, it suggests shorter conversations, summarized documents, clearer first requests, smaller models for simple tasks and piloting workflows with budgets and auto-refill turned off.
Perplexity explains harness engineering, the design of everything around an AI model that makes agents reliable.
Perplexity defines harness engineering as designing every component of an AI tool apart from the model, including how information reaches the model and how its work is checked.
Because a model alone handles only text and forgets earlier exchanges, the harness supplies history, tool access and oversight.
Harnesses matter most for agents performing multi-step tasks, where an early mistake can affect every later step.
Perplexity says agent harnesses typically include an agent loop, context management, tools, memory, permissions and verification.