Skip to content
/
Phase B Batch 18

Batch 18 — Sandboxes + Execution Isolation

Batch 18 — Sandboxes + Execution Isolation

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
brood-box 36 standalone-repo (Go) bbox (yes) none none (1 agent/session) no A
vibebox unknown unknown unknown unknown unknown unknown C
daytona 72,455 standalone-repo (multi) daytona (yes) web-dashboard parallel-fan-out no A
e2b 12,361 npm-package + pip e2b (yes) none parallel-fan-out no A
microsandbox 6,299 standalone-repo (Rust) msb (yes) none parallel-fan-out no A
agent-sandbox-k8s 2,453 standalone-repo (k8s CRD) none (kubectl) none none no A
arrakis 816 standalone-repo (Go) arrakis-client (yes) vnc none no A
vibekit 1,790 npm-package vibekit (yes) unknown none yes A
flydex unknown unknown unknown unknown unknown unknown C
worktrunk 5,197 cli-tool (Rust) wt (yes) terminal-tui parallel-fan-out no A

Intra-batch patterns

All 8 analyzable frameworks in this batch sit below the agent loop rather than inside it — they provide execution substrates (microVMs, containers, Kubernetes pods, git worktrees) rather than behavioral instructions (skills, hooks, personas). This is the defining characteristic that separates them from all 11 seed frameworks. Five of the eight (brood-box, microsandbox, arrakis, and the two Tier C entries) use hardware microVM isolation via libkrun or cloud-hypervisor, making hardware virtualization the dominant isolation approach in this batch. Two use cloud/container isolation (E2B, Daytona) and one uses Kubernetes container primitives (agent-sandbox-k8s). Worktrunk is the outlier — it uses git-worktree isolation (matching seeds like superpowers) and is the only framework in the batch that injects any agent-facing content at all (a Claude Code skill for configuration guidance + activity-tracking hooks). The batch reveals a clear market segmentation: local microVM (brood-box, microsandbox, arrakis), cloud API (E2B, Daytona), Kubernetes operator (agent-sandbox-k8s), agent routing layer (vibekit), and worktree management (worktrunk).

Most interesting find

Arrakis: VM-level snapshot-and-restore as a first-class primitive for AI agent backtracking — enabling MCTS-style exploration patterns that no other framework in the batch (or seeds) supports. The snapshot captures full VM state including running processes, making it possible to restore an agent to the exact computational moment of the snapshot.

Worktrunk: The only framework in this batch that ships both a Claude Code plugin (with BeforeAgent/AfterAgent/SessionEnd hooks for activity tracking) and a meaningful skill file — bridging the gap between "worktree infrastructure tool" and "agent-aware workflow tool." The activity marker system (🤖/💬 per worktree in wt list) is a novel observability primitive not seen in the seed frameworks.

Items written as Tier C

  1. vibebox — No public GitHub repository. URL is a product landing page (vibebox.robcholz.com). Cannot analyze without web scraping (prohibited under research constraints).
  2. flydex — No public GitHub repository. URL is a product landing page (flydex.net). Cannot analyze without web scraping (prohibited under research constraints).

Cross-references discovered

  • VibeKit (vibekit) directly depends on E2B (@vibe-kit/e2b) and Daytona (@vibe-kit/daytona) as pluggable sandbox provider packages — three frameworks in this batch are in a direct dependency relationship.
  • Brood Box and Microsandbox both use libkrun as their microVM backend (from the same containers/libkrun project) — they are architectural siblings using the same virtualization layer.
  • Arrakis uses cloud-hypervisor as its VMM, which is a different (higher-feature) alternative to libkrun — same hardware isolation goal, different implementation.
  • Worktrunk (wt) is explicitly cited in Anthropic's "Claude Code Best Practices" guide for the worktree pattern, giving it a quasi-official endorsement for the parallel agent workflow.
  • Microsandbox ships superradcompany/skills (separate repo) to teach Claude Code / Cursor / Codex / Gemini CLI / Copilot how to use it — the skills bridge this infrastructure tool to the agent instruction layer, creating a dependency chain from microsandbox (this batch) → skills framework (seed archetype 1).