Perplexity disclosed in its official engineering blog on October 12, 2026, that Projects, the feature already live in Perplexity Computer, does not sit on an ordinary shared folder. Underneath is a Git repository system built specifically for collaboration between people and agents: every Project maps to a single Git repository, and a user's interactive agents, scheduled agents, and other collaborators' agents all commit their changes into that same repository. No new model was announced, but the post addresses a problem that becomes unavoidable once agents start working over long horizons: where the work lives, who changed what, and how to roll back when something goes wrong.
Why Git, of all things
Perplexity's reasoning is practical. Collaboration that stretches over months needs three properties: the work must persist across sessions instead of vanishing when a sandbox is destroyed; simultaneous edits by multiple people and agents must not silently overwrite each other; and every change must be traceable and reversible. Those are exactly the problems Git has solved for decades. The difference is that traditional code forges were designed for humans, and their review screens, branch management, and notifications are dead weight for agents. Perplexity therefore keeps only Git's storage and verification machinery. Agents do not wait for human approval; instead, the Git service mechanically checks each submission against a repository contract and accepts it as the new head only if it passes. When two agents race, the later submission built on a stale base receives an explicit rejection; that agent replays its changes on top of the new head and submits again. In production, Perplexity says, fewer than 2% of submissions are rejected as stale, median submit latency is under 1.7 seconds, and the 99th percentile is under 8 seconds.
How the five parts divide the work
The system has five components. Each agent works in its own isolated sandbox, where a small command-line tool materializes the repository locally and submits work in meaningful units. The control plane only locates repositories and issues short-lived credentials; it never touches repository data. The Git service runs Git's native clone, fetch, and receive operations and decides which submission becomes the latest accepted commit, with state recorded in DynamoDB. Repository data itself lives on a shared filesystem provided by AWS S3Files, so service replicas stay stateless and replaceable. After a commit is accepted, background workers extract text and metadata from changed documents so agents can search them the next time a project is opened. To save bandwidth, sandboxes do not download whole repositories up front: large files first appear as Git LFS pointers, agents read the extracted text, and originals are fetched only when necessary. Perplexity notes that it tried a FUSE-based virtual filesystem and abandoned it after finding it roughly 400 times slower on file-heavy workloads. History does not grow without bound either; once retained commits pass a threshold, older history is compacted into a single starting snapshot.
What the numbers show, and where the claims stop
The internal benchmark Perplexity published illustrates the design's intent. On 150 financial questions over the same document corpus, an exploration agent working through files scored 0.867 at retrieving the correct document and 0.627 at answering correctly, against roughly 0.39 and 0.37 for semantic search — about 1.7 times better on answers, at about 25% lower latency. Reattaching to a warm project checkout takes about 2.9 seconds, versus roughly 8 seconds cold. Computer's long-term memory, Brain, runs on the same repository infrastructure, with knowledge pages stored as commits in separate repositories so memory updates never compete with project edits. Two cautions apply: every figure is self-reported by Perplexity with no third-party replication, and the test corpus was of its own choosing, so the numbers document a design rationale rather than an industry verdict. The direction still matters. As agents move from one-shot answers to collaboration measured in months, competition is shifting from the model itself to the persistence and auditability of the workspace. Perplexity's bet is that version control, invented decades ago for human collaborators, is the fastest foundation agents can borrow — and whoever stabilizes that layer first will have agents trusted with longer work.
