Top AI Product

Coding agents have one nasty habit: they nuke your working directory. Half-applied edits, deleted files, an rm -rf nobody asked for. Tilde.run hit 162 points on Show HN today by treating that as a database problem — every agent run is a transaction. Clean exit commits, crash rolls back, nothing silently overwritten.

What it actually is

A sandboxed POSIX filesystem agents mount as ~/sandbox. It pulls code from GitHub, training data from S3, and documents from Google Drive into one versioned tree. Every file is versioned from the first write, every outbound network call logged. Three guarantees the team keeps repeating: reversibility, isolation, audit. Built on top of lakeFS — the data versioning layer Treeverse has been shipping since 2020 — so the engine has production miles.

API for agent runtimes

The sandbox API is built for agent frameworks, not humans. Your runtime opens a session, the agent does whatever destructive thing it wants, you review the diff, commit or discard. Think git stash if the stash were a fully isolated container covering GitHub, S3, and Drive at once. For anyone running Claude Code, Cursor, or a custom agent loop on real production data, this is the missing safety layer — stop praying the agent doesn’t go rogue, run it like a transaction.

You Might Also Like

• Google A2ui Agent to User Interface Finally a Standard way for ai Agents to Show you Things
• Emdash Lets you run 21 ai Coding Agents at Once and it Actually Works
• Superset Just Made Running a Whole Squad of ai Coding Agents Actually Manageable
• Memento Wants to Record Your ai Coding Sessions Right Into git and its Sparking a Real Debate
• Agent Action Protocol aap the Missing Layer Above mcp That Actually Makes Agents Production Ready