Returning to Development States: When Full Sandbox Snapshots Help Coding Agents
The paper argues that snapshotting the whole sandbox helps most when useful progress lives outside the code diff.
Graviet and Shah test a Go-Explore-style recovery setup that restores full Daytona environment snapshots instead of restarting failed coding-agent runs from scratch. The reported gains come mainly when intermediate state includes databases, running services, or other runtime context that a repository patch would not capture. The authors frame this as most relevant for longer coding tasks spanning multiple services, where rebuilding state from files alone is unreliable. Preprint, not peer reviewed. ArXiv · AI/CL/LG's note
Graviet and Shah test a Go-Explore-style recovery setup that restores full Daytona environment snapshots instead of restarting failed coding-agent runs from scratch. The reported gains come mainly when intermediate state includes databases, running services, or other runtime context that a repository patch would not capture. The authors frame this as most relevant for longer coding tasks spanning multiple services, where rebuilding state from files alone is unreliable. Preprint, not peer reviewed. ArXiv · AI/CL/LG's note
score 5