50
Decentralized swarms meet OS-level sandboxes under the agent loop
Power users get clone-and-run multi-agent coordination that scales without a central boss plus lightweight isolation primitives that keep credentials and host state out of every coding session. The through-line is control planes and sandboxes you wire yourself rather than closed IDEs.
01 — Power tools
The stack
- 01harnessGitHubAgensh↗
Microsoft Research (Zhihao Zhan et al.)
Self-organized multi-agent harness that scales to 1024 agents with async task claiming, shared workspace, and no central orchestrator.
Use caseUse it to spin up large peer teams that rebuild reference programs from observations on ProgramBench-style tasks.
multi-agentdecentralizedresearch-origin
02 — In the wild
Articles and releases
- releaseemergent.shAnthropic open-sources sandbox-runtime for container-free agent isolation↗
Anthropic / Emergent
srt ships as an OS-level tool using Seatbelt on macOS and bubblewrap on Linux plus network proxies. It powers Claude Code security and is now available for any agent or MCP server that needs filesystem and egress restrictions without containers.
sandboxreleasesecurityShipped
- articleinfrazen.ioAI Agent Sandboxes Compared: eight options platform teams are evaluating this quarter↗
InfraZen Engineering
The piece ranks gVisor, Kata, Firecracker and policy sandboxes by isolation boundary, maturity and the ten controls that actually stop recent agent escapes. It argues most published failures were control gaps rather than kernel escapes and maps each option to laptop versus cluster use.
comparisonsandboxengineeringRead