Agentic AI
Perplexity's local agent splits reasoning from execution control to boost enterprise security
Separating what models propose from what code actually permits is becoming a core engineering challenge for agents running in enterprise environments.
The New Stack
Perplexity released Portable Computer this week—a locally-hosted version of its agent that runs on Nvidia DGX Spark hardware—marking a significant architectural shift in how enterprise-grade AI systems handle reasoning versus authority.
The infrastructure is expensive: a DGX Spark starts at $4,700, with older GPUs like the 24GB RTX 3090 commanding well over $1,500 secondhand. But the real story lies in how Perplexity engineered the agent's decision-making pipeline.
A different approach to agent reliability#
Rather than stacking multiple models to oversee one another—the standard approach of adding orchestrator, planner, or critic models to improve reliability—Perplexity created a two-tier architecture. The local model proposes actions, including which tools to call or when to seek cloud assistance. Deterministic code then decides whether those proposals get executed. This mirrors control-plane thinking: reasoning suggests, but inspectable software retains the final say.
The runtime controller assembles context, enforces policy constraints, and runs approved tool calls within an OS-level sandbox. According to Nate Kupp, Perplexity's vice president of Computer Enterprise and Infrastructure, this separation meant the execution harness accounted for the bulk of engineering effort.
Same model, better outcomes through system design#
Perplexity kept Qwen 3.8-27B and the same hardware constant while testing three agent stacks on its internal Local Knowledge Work Bench—a held-out set of 53 real-world tasks. Computer achieved 82.6% accuracy, outpacing its Pi agent at 77.6% and Hermes at 74%. The performance gains weren't from a larger or smarter base model; they came from the system wrapping it.
On ParseBench-100—which tests chart reading, page layouts, tables, and formatting—the gap widened dramatically: Computer hit 65.1% versus Hermes' 34.6% and Pi's 13.9%. This disparity demonstrates how orchestration, prompting, tool schemas, and context management matter as much as raw model weights.
Perplexity also post-trained Qwen into a variant called PPLX 27B, which scored 85.4%—showing that harness engineering and fine-tuning work together. The "harness" encompasses prompts, tool schemas, verification logic, and document processing pipelines, not just the execution layer.
