Skip to content
/
Phase D Batch 30

Batch 30 — Cloud Sandbox Runtimes & Browser/Desktop Harness Platforms

Batch 30 — Cloud Sandbox Runtimes & Browser/Desktop Harness Platforms

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
ms-agent-framework 10,763 standalone-repo (PyPI + NuGet) devui (web app launcher) yes (web, port 8080) parallel-fan-out (graph) yes A
agentscope-runtime 800 standalone-repo (PyPI) no yes (starter WebUI) other (HTTP/SSE) yes A (archived)
e2b-desktop 1,388 standalone-repo (PyPI + npm) no no none no A
agentbay-sdk 1,124 standalone-repo (multi-lang) no no none no A
tensorlake 926 standalone-repo (PyPI + CLI) tensorlake no parallel-fan-out yes A
openshell-nvidia 6,264 standalone-repo (binary) openshell yes (terminal-tui) sequential yes A
agent-infra-sandbox 4,809 docker-image unknown yes (web, port 8080) none no A
browser-harness 13,821 standalone-repo (PyPI) browser-harness no parallel-fan-out no A
kubestellar-kc-agent 109 standalone-repo (Go binary) kc-agent yes (web, port 8080) none yes A
swarmvault 492 npm-package swarmvault yes (web + desktop) sequential yes A

Intra-batch Patterns

This batch splits cleanly into three sub-groups: (1) Cloud sandbox infrastructure (E2B Desktop, AgentBay SDK, Tensorlake, AgentScope Runtime) — all disposable VM/container APIs with no pre-authored agent behaviors, no CLI for end users (except Tensorlake), and "bring your own agent" philosophy; (2) Local/self-hosted sandbox runtimes (OpenShell, agent-infra-sandbox) — Docker or policy-governed container/MicroVM runtimes where the framework ships with behavioral conventions (OpenShell has 19 skills + 2 Claude subagents, agent-infra ships MCP-native all-in-one container); and (3) Browser + knowledge harnesses (Browser Harness, SwarmVault, KubeStellar Console) — all target Claude Code/Codex as primary agents and ship skill-md files or CLAUDE.md guidance. The Microsoft Agent Framework spans categories — it is the only non-Python production multi-agent SDK with a DevUI dashboard, and the only one targeting .NET. A striking pattern: 6 of 10 frameworks ship AGENTS.md and/or CLAUDE.md files indicating they use or document Claude Code for their own development.

Most Interesting Finds

OpenShell (NVIDIA): The most architecturally sophisticated framework in this batch. Combines a full Rust runtime with four-layer policy enforcement (filesystem/network/process/inference with hot-reload), a 19-skill agent development workflow, two Claude Code subagents (arch-doc-writer on opus, principal-engineer-reviewer), and the most principled human-gating design in the corpus (state:agent-ready is documented as "non-negotiable safety control" that agents must never bypass). The observation that "OpenShell is built agent-first — we design systems and use agents to implement them, this is not vibe coding" reflects a mature philosophy.

Browser Harness: The self-improving harness pattern — where the agent writes missing capabilities to agent_helpers.py during execution — is the most novel execution paradigm in the batch. Combined with the principle that domain skills should be agent-generated (not hand-authored), it represents a fundamentally different approach to tool extension compared to all other frameworks.

Items Written as Tier C

None. All 10 frameworks had sufficient public material for full 11-file reports.

Note on AgentScope Runtime: Archived/transitioning to AgentScope 2.0. Written as Tier A because material was complete, but flagged as archived in METRICS.yaml (maintainer_status: archived).

Cross-References Discovered

  • E2B Desktop is the desktop surface of the canonical e2b seed from Phase B Batch 18/33; explicitly an extension of e2b-dev/desktop on top of the E2B Sandbox MicroVM platform.
  • AgentBay SDK is closest to E2B Desktop in pattern (disposable cloud sandbox API) but from Alibaba Cloud (Wuying infrastructure). The SDK's hooks/ directory and MCP mentions suggest future convergence with MCP.
  • Tensorlake explicitly benchmarks against E2B, Modal, Vercel, and Daytona (all in Phase B canonical sandbox batch), positioning itself as the performance leader.
  • AgentScope Runtime documents Microsoft Agent Framework as a supported framework adapter, creating a direct dependency relationship between two items in this batch.
  • OpenShell references NemoClaw (NVIDIA's OpenClaw runtime) and the OpenShell-Community repo for sandbox images — the community catalog includes Claude Code, Codex, OpenCode, GitHub Copilot as pre-installed agent containers.
  • SwarmVault credits Andrej Karpathy's LLM Wiki gist as origin pattern. The Obsidian plugin in packages/ suggests the ccmemory cross-pattern (both build on graph-based knowledge for agents).
  • Browser Harness was created by browser-use.com, same team as the Browser Use library. The harness and library are separate products serving different niches (harness = user's real Chrome; library = headless automation).
  • KubeStellar Console references Kagenti (separate CNCF-adjacent project) as a backend integration, suggesting it is the frontend for a broader multi-project ecosystem.