What happened
Perplexity used Computex 2026 to show a hybrid local-cloud inference orchestrator for its Personal Computer agent. The claim is not simply that smaller models can run locally. The more interesting part is task-level routing: the system decides which parts of a workflow stay on the device and which parts are sent to frontier models in the cloud.
In the demonstration, the agent processed confidential deal materials and kept sensitive information local while routing heavier reasoning work to cloud models when appropriate. VentureBeat reports that the feature is not yet generally available, but Perplexity says it should launch in the coming weeks.
This fits a broader shift in enterprise AI. Agents are moving beyond chat windows into file systems, business applications, spreadsheets, SharePoint, CRM and workflow tools. Once agents touch operational data, the question changes from which model is smartest to where each step runs, who can inspect it, and how the organization proves what happened.
Why it matters
Hybrid inference is becoming a practical architecture pattern for enterprise AI. Fully cloud-based agents are easy to start with, but they create real concerns around data residency, confidential documents, token cost, latency and vendor dependency. Fully local systems offer more control, but they can struggle with complex reasoning, model updates and operational support.
A routed runtime tries to avoid that false choice. Sensitive extraction, classification or summarization can happen near the data. Less sensitive reasoning, enrichment or synthesis can use external models when they add value. That model-agnostic split is exactly where the enterprise conversation is heading.
There is also a cost angle. Agent workloads are not one prompt and one answer. They can run multi-step plans, call tools, inspect documents, retry, verify and log. If every intermediate step goes to premium cloud models, costs become hard to forecast. If every step runs locally, quality may suffer. Routing lets teams reserve expensive inference for the parts that need it.
Laava perspective
For Laava, the important lesson is that the runtime is becoming the product boundary. Customers do not need another loose AI tool or a hardware box with a logo on it. They need a managed environment where agents can work with documents and systems under clear rules.
That is why Laava frames Sovereign Runtime and Laava Box as deployment forms inside Laava Agents and Custom Solutions. The value is not local compute by itself. The value is managed runtime, agents, integrations, monitoring, logging, updates and governance around real operational work.
Perplexity’s announcement reinforces the same direction: model choice, location choice and auditability belong in the architecture. A useful enterprise agent should be able to use Azure OpenAI today, an open model tomorrow, and local inference when the data or economics demand it. The customer should not have to rebuild the workflow every time the model market changes.
The hard part is not the slogan “local plus cloud”. The hard part is designing the routing policies, permissions, logging, fallback behavior and integrations so the system is reliable in production. That is engineering work, not AI theatre.
What you can do
If you are exploring AI agents, start by mapping which data can leave your environment, which data should stay close, and which actions require a human approval trail. That map is more useful than a model benchmark when deciding what architecture you need.
Then pilot one workflow where control matters: contract review, ticket triage, SharePoint knowledge search, email handling or internal reporting. Prove the runtime, permissions and logs before scaling. The companies that win with agents will not be the ones with the flashiest demo, but the ones that can explain exactly where the work ran and why.