The boundary between tools that merely answer questions and systems that actually perform work has been dissolving for some time, yet the practical consequences of that dissolution are only now becoming concrete.
For years the dominant pattern remained conversational: a user posed a query, received a response, and then carried the output into separate applications to act on it.
That pattern is beginning to invert.
Agents are acquiring the ability to operate directly on the files, applications, and workflows that already constitute daily labor, and the organizations deploying them are simultaneously confronting the need for visibility and control at the same layer where the work occurs.
Against this backdrop, a cluster of releases from Perplexity in late July 2026 illustrates how one company is attempting to close the remaining gaps between capability, integration, and accountability.
First off, Personal Computer, now available inside the Perplexity application for Windows 10 and 11, functions as a local agent harness rather than a remote chat interface.
Once launched, it can be pointed at approved folders on the desktop and thereafter opens, reads, and edits the files inside those folders without requiring the user to shuttle content between windows.
It routes individual subtasks across a roster of more than fifteen models, selecting specialists for structured extraction, real-time web grounding, long-context synthesis, or high-stakes judgment according to the nature of the work.
Persistent memory, branded Brain, supplies each new task with the accumulated context of prior sessions, projects, and files so that continuity is maintained without repeated explanation.
The same system reaches outward to more than 400 connected applications, including Salesforce, Snowflake, HubSpot, and the Microsoft 365 suite, pulling data into local files and operating natively inside Excel, PowerPoint, Word, and Outlook.
Voice input is supported, as is direct control of the Comet browser for research and web automation.
The result is an environment in which research, coding, document generation, and administrative triage can unfold inside a single continuous session rather than a fragmented sequence of copy-paste operations.
Layered onto that foundation is Model Council.
When a user elects this mode, they assemble a temporary panel of two to eight frontier models drawn from OpenAI, Anthropic, Google, and open-source providers, then choose the depth of analysis required.
Each model examines the question independently; the system returns a single cited report that distinguishes regions of consensus from points of disagreement and surfaces observations unique to particular participants.
Because the council operates inside Computer, the synthesis is not left as abstract prose.
It can be converted on the spot into board decks, legal memos, valuation spreadsheets, or other work-ready artifacts.
The feature is especially suited to ambiguous or high-stakes decisions where no single model’s judgment is presumed sufficient: contract risk ranking, pre-revenue valuation, build-versus-buy analysis under tight timelines, or personal financial trade-offs.
By forcing models to declare their assumptions and confidence levels, the council converts divergence itself into useful signal rather than unresolved noise.
The third element addresses the security implications that arise once agents gain the ability to act autonomously on privileged endpoints. Numbat, released as open source under the Apache 2.0 license, is an agent-detection and response layer designed to operate across desktop, CLI, IDE, and gateway harnesses.
Numbat integrates through hooks, session artifacts, and OpenTelemetry telemetry, normalizing activity from heterogeneous agents into a common event model.
A local CEL rule engine evaluates those events in real time, covering categories such as privilege escalation, secret access followed by egress, unauthorized file modification, and multi-step attack sequences.
Pre-action hooks allow selected rules to block an operation before it executes; all shipped rules begin in monitor-only mode so that enforcement remains an explicit operator choice. Forensic reconstruction is possible even for sessions that predate Numbat’s installation, because the tool can parse existing on-disk artifacts into normalized timelines.
Inside Perplexity the same system feeds audit events back into Computer, which analyzes the telemetry, escalates anomalies, and proposes refined local detections, creating a continuous improvement loop.
Distributed as a single static Go binary for macOS, Linux, and Windows, Numbat supplies security teams with a consistent interface rather than forcing them to instrument each agent framework separately.
Taken together, the three releases sketch a coherent architecture.
Agents are no longer confined to answering questions or generating text; they operate on the local filesystem, across enterprise applications, and through multi-model deliberation.
At the same time, their actions become observable, policy-governable, and retrospectively auditable at the harness level rather than solely inside the model.
The practical effect is to move autonomous assistance from an experimental overlay into the same operational and security regime that already governs other privileged software running on enterprise endpoints. This implementation is less important than what it demonstrates: the technical foundations for this kind of integration are now coming together.
Local orchestration, multi-model synthesis, and system-level controls are being assembled, while the security layer has already been released as open source.
