Perplexity Explains Projects: A Git-Based Shared File System for Agents
Perplexity published a long official architecture thread, offering the first systematic look at the design behind Projects: building a Git-style shared filesystem for multiple agents so their work can persist across sessions and agents.
Confirmed
- Core problem: when multiple agents collaborate on long-running projects, they face the same issues engineers have with shared codebases—work must outlive ephemeral sandboxes, concurrent writes must not clobber each other, and history must trace who changed what (@perplexityai)
- Solution: each agent works in its own sandbox and commits to a shared Git repository; humans and other agents can build on those results across sessions
- Sandbox strategy: long-lived sandboxes were dropped (they must be provisioned for peak load and kept online for months—costly and fragile); instead, the repository sits on S3-backed storage and agents run in ephemeral sandboxes—a crash only loses uncommitted work, and the Git service itself is unaffected
- Concurrency control: the CLI holds a short-lived lock to serialize commits, and the Git service only accepts submits based on the latest head; otherwise the agent replays its changes on the new head and resubmits
- Internal eval: on 150 financial-document questions, agents exploring a real checkout with grep and other standard file tools hit 0.627 answer accuracy versus roughly 0.37 for semantic search—nearly double for the exploratory approach; it also scored better on retrieval
Why it matters
- The design offers a reusable engineering paradigm for long-horizon multi-agent collaboration: solving agent state persistence and concurrent writes with battle-tested Git semantics (atomic multi-file commits, traceable history, conflict handling) rather than fragile long-running sandboxes
- The finding that grep-style exploration significantly beats semantic search is directly relevant to retrieval tool choices for agents
2026-10-13 ~ 2026-10-13 · 6 related posts
Primary sources
- Perplexity built Projects' filesystem as a Git forge for agents, with greps beating semantic search — perplexity_ai ·
- Perplexity frames long-running agent projects as a shared-codebase problem — perplexity_ai ·
- Perplexity: agent grep exploration hits 0.627 answer accuracy vs ~0.37 for semantic search — perplexity_ai ·
- [source] Perplexity built Projects' filesystem as a Git forge for agents, with greps beating semantic search — perplexity_ai · 2026-10-13
- [source] Perplexity frames long-running agent projects as a shared-codebase problem — perplexity_ai · 2026-10-13
- How Perplexity serializes concurrent agent submits with head checks and replay — perplexity_ai · 2026-10-13
- Perplexity swaps long-lived sandboxes for ephemeral ones on S3-backed storage — perplexity_ai · 2026-10-13
- [source] Perplexity: agent grep exploration hits 0.627 answer accuracy vs ~0.37 for semantic search — perplexity_ai · 2026-10-13
1 near-duplicate retellings: perplexity_ai