Perplexity grows up: from answer engine to research agent.
Perplexity started as a faster, cited way to search. In June 2026 it kept moving toward something bigger: a research agent that runs multi-step investigations for you inside Perplexity Computer. New this month — Deep Research running in Computer, a command panel, and forking. Here's what each actually does, what changed under the hood, and the one habit that keeps agentic research from quietly leading you astray.
01 The shift: search → investigate
A normal search answers one question. Deep Research takes a goal — "compare these five vendors on price, security, and support" — and runs a whole sequence of searches, reads sources, and assembles a structured, cited report. The June update is that this now runs inside Perplexity Computer, the agent workspace, so a long investigation happens in one place instead of a string of one-off queries.
It's not magic — it's delegated legwork. The agent does the dozens of searches and the first-pass reading you'd otherwise grind through yourself, then hands back something organized with its sources attached. The thinking about whether the conclusion is right is still yours. That division of labor is the whole point — and the whole risk.
02 What's actually new this month
| Addition | What it does |
|---|---|
| Deep Research in Computer | Runs full multi-step research inside the Computer agent workspace |
| Command panel | A faster way to steer and issue commands to the agent mid-task |
| Forking | Branch a research thread to explore an alternative without losing the original |
| Stronger model under Deep Research | Perplexity Deep Research now runs on a more capable reasoning model, improving results on its benchmarks |
| Enterprise controls | Computer Analytics API and custom credit limits for teams |
Forking is the quietly useful one. Research rarely goes in a straight line — you want to chase a tangent without torching your main thread. Forking lets you branch, explore, and compare instead of starting over.
You're deep in a vendor comparison and want to explore "what if we only cared about security?" without losing your current report. What's the right move?
03 The habit that matters: read the citations
Perplexity's whole advantage over a plain chatbot is that it shows its sources. Agentic research makes that more important, not less — because the agent did a lot of reading fast, and a confident, well-formatted report can hide a weak source or a misread.
This is exactly why a tool that exposes its sources is worth more than one that doesn't: it makes verification possible. But possible isn't automatic — you still have to do it on the claims that count.
04 Hybrid compute shipped — on Mac, September 1, 2026
The hybrid local/cloud inference Perplexity previewed at Computex is no longer a preview. Since September 1, 2026 the Perplexity for Mac app lets Computer split a task between a compact local model running on your Mac and the frontier models in the cloud. A privacy gate on the machine decides what may leave it: before anything from a protected file reaches the cloud it can mask the sensitive details, keep the step local, refuse, or ask you first. The cloud still does the heavy reasoning, web search and planning; the local model handles the private files and on-device actions.
What you need: a Perplexity Pro, Max or Enterprise subscription, an Apple silicon Mac on macOS 15+ with at least 24 GB of unified memory, and one of the three launch local models (Gemma 4 E4B, Qwen3.6 35B-A3B, or Perplexity's own). Download a model from the Mac app, open the model selector, pick hybrid, choose the local and cloud models — done. Enterprise admins can set org-wide rules for what must stay on the Mac and audit when information leaves a device. Windows users: Personal Computer runs on Windows, but hybrid compute is Mac-only at launch.
05 What should you use it for?
Your move
Run one real Deep Research task in Computer, then do something most people skip: pick the three claims the conclusion most depends on and click through to their sources. Note whether each source actually says what the report claims. That five-minute check is the difference between using a research agent well and being quietly misled by a confident summary.
More Perplexity lessons →What you can do now
- Use Deep Research in Computer for multi-source investigations, not one-off facts
- Fork a thread to chase a tangent without losing your main report
- Treat agentic research as delegated legwork — the judgment on the conclusion stays yours
- Always spot-check the load-bearing citations; a cited answer isn't automatically a correct one
- Hybrid local/cloud compute shipped on Mac on September 1, 2026 (Pro/Max/Enterprise; Apple silicon, 24 GB+); Windows is still cloud-only