Perplexity Details Git Based Workspaces for AI Agents

Perplexity AI detailed in a company blog post how Projects gives people and AI agents persistent shared workspaces for long running tasks. Each project uses a Git repository to preserve files across sessions, manage concurrent edits and record who changed each file.

Agents work in separate temporary sandboxes and submit completed work as commits through a command line tool. A control plane locates each repository and issues temporary credentials, while a Git service accepts submissions and stores repository data through AWS S3Files. If two agents submit changes from the same repository state, the service accepts one and asks the other to replay its work on the new version.

Perplexity said fewer than 2% of production submissions receive this stale base response. Median submission latency is below 1.7 seconds, while p99 latency is below 8 seconds. Background workers also extract text and metadata from documents, allowing agents to search PDFs and Office files with standard file tools without downloading every original file.