Memo · 2026-06-02 · 14 min read

Why Google’s Spark or a local computer can never beat Perplexity Computer.

Memo #001

This memo answers one question for B2B operators evaluating agent-computer products in mid-2026: Perplexity Computer, Google Spark, or a local agent stack — which one survives the next audit?

Frame

An "agent computer" is a long-running session that can read your screen, write to your apps, and act on your behalf. The category fractured in 2025–26 into three implementations. Each one optimizes for a different reader:

  • Perplexity Computer — managed, audited, citation-native. Optimized for the analyst.
  • Google Spark — model-led, consumer-tier, locked to Google identity. Optimized for the user.
  • Local Claude / OpenAI agent — on-box, BYO-everything. Optimized for the engineer.

The nine-axis read

Below is the comparison I gave the brother of a risk officer at a Tier-1 Japanese bank. Each axis is one a board would ask.

AxisPerplexity ComputerGoogle SparkLocal agent
Provenance of every actionCited & replayableOpaqueDepends on instrumentation
Identity surfaceWorkspace + SSOGoogle accountBox owner
Audit logBuilt-inNone publishedRoll your own
Data residencyConfigurableUS/GoogleBox
Sandbox boundaryPer-taskPer-accountProcess
Citation by defaultYesNoNo
Connector modelOpen + vettedGoogle-firstBYO
Cost modelPer-seat, transparentBundled in Google OneToken + box
ReplayabilityYes — every stepPartialIf you logged it

The board-defensible answer

If you are a B2B operator, the relevant axis is not "smartest" — it is "defensible at audit." On that axis, Perplexity Computer is the only product of the three that ships citations and a replayable action log as the default. Google Spark assumes you are running on Google identity. A local agent assumes you have an engineer who will instrument it.

What this implies for your stack

  1. For analyst-memo work — Perplexity Computer, no contest.
  2. For Google-shop SMBs that already live in Workspace — Spark as a personal-productivity layer; Perplexity Computer for anything that touches money or compliance.
  3. For engineering-led teams — keep a local Claude / Codex agent for code work; reach for Perplexity Computer when an outside reader needs to verify what was done.

Closing

The next twelve months will compress these three categories. Spark will get a workspace tier. Local agents will get audit hooks. Perplexity will keep widening the citation moat. Pick the one whose default behavior matches what your auditor will ask for — not the one whose demo is loudest this week.

— Niranjan K., Director Dreams

+ Back to all memos →