36
MicroVM control planes and cloud-native harnesses outpace chat wrappers
Today's picks center on programmable isolation layers and replaceable agent loops that power users can clone, wire, and run under their own policies. Sandbox-as-tool patterns and durable session control planes dominate over single-vendor IDEs.
01 — Power tools
The stack
- 01repoGitHubGondolin↗
earendil-works
Local Linux micro-VMs with a TypeScript control plane for programmable network, filesystem, and secret injection policies.
Use caseUse it to wrap agent tool calls in disposable QEMU or krun VMs while keeping credentials and egress on the host.
microvmsandboxtypescriptcontrol-plane - 02harnessGitHubMecatl↗
stacklok
Cloud-native open-source agent harness with provider-agnostic loop, durable state, permissions, gRPC/HTTP APIs, and optional local microVM execution.
Use caseUse it to run the same agent loop locally or on Kubernetes with Redis-backed sessions and deny-dominant permissions.
harnesskubernetescloud-nativepermissions - 03repoGitHubzerocore-ai/microsandbox↗
zerocore-ai
Open-source local-first hardware-isolated sandboxes with instant startup and OCI compatibility for agent workloads.
Use caseUse it as an embeddable microVM backbone for untrusted code execution in agent pipelines.
microvmlocaloci - 04toolOrigins AI / productPrime Sandboxes↗
Prime Intellect
Hardware-isolated Linux microVMs for agentic RL at scale with large immutable environment catalog and high concurrency.
Use caseUse it to train or evaluate agents against reproducible microVM environments at thousands of concurrent instances.
microvmrlevalcatalog - 05harnessnpm@ai-sdk/harness↗
Vercel AI SDK
Experimental HarnessAgent plus sandbox session interfaces for network-aware agent loops in the AI SDK.
Use caseUse it to bootstrap typed harness sessions with Vercel or E2B sandboxes and schema-backed outputs.
sdkharnesstypescriptsandbox - 06toolGitHub MarketplaceAgent Security Harness↗
msaleme
CI action that runs protocol-level adversarial tests against MCP/A2A endpoints with PASS/FAIL/INCONCLUSIVE verdicts.
Use caseUse it as a gate to probe live agent endpoints for policy violations without false PASS on unreachable targets.
securityevalcimcp - 07repoGitHube2b-dev/ai-sdk-sandbox↗
e2b-dev
E2B sandbox backend for Vercel AI SDK so HarnessAgent code runs inside microVMs.
Use caseUse it to wire E2B sessions into AI SDK tools and harnesses with restricted surfaces.
e2bai-sdksandbox - 08repoGitHubRyanAlberts comparisons↗
RyanAlberts
Category-ranked top harnesses regenerated weekly from live star and capability data.
Use caseUse it to shortlist progressive-disclosure, coding, or research harnesses by current metrics.
rankingcomparisonharness
02 — In the wild
Articles and releases
- releaseAWS Open Source BlogAWS launches Strands Box open-source agent sandbox↗
Fernando Dingler / AWS
Strands Box combines OS isolation with Dogwood policies, Strands Shell, and Monty so agents face fine-grained action controls rather than YOLO mode. It is Apache-2.0, macOS-first, and works with any harness.
awssandboxreleasepolicyShipped
- articleCNCFMecatl: the case for a cloud-native agent harness↗
Craig McLuckie
Stacklok argues agent runtimes must be distributed systems with durable state, leases, and replaceable execution planes. Mecatl ships as the open reference that keeps the loop independent of clients and infrastructure.
cncfharnesskubernetesessayRead