48
Local microVMs and open harnesses take the agent loop offline
Power users are cloning branchable hardware-isolated sandboxes and thin control planes, wiring them into durable loops instead of closed IDE chatbots. Isolation, checkpoints, and programmable runtimes now sit under the model rather than around a vendor UI.
01 — Power tools
The stack
- 01harnessGitHubagent-eval-harness↗
Siddharth-1001
Local-first evaluation framework that traces full agent behavior, not just LLM outputs, with structured metrics and side-by-side comparisons.
Use caseUse it to measure tool success rates, cost creep, and regressions without sending traces to a hosted platform.
evalharnesslocaltraces - 02toolGitHubVmette↗
chamuka-inc
Hardware-isolated microVM sandbox aimed at local AI agents on macOS, recently shown on HN for safe full-permission runs.
Use caseUse it to run agents with root-like freedom inside a guest while the host stays untouched.
microvmmacossandboxlocal
02 — In the wild
Articles and releases
- releaseAI Agent Store / TanStackTanStack AI 1.0 stabilizes agent toolkit and experimental harnesses↗
TanStack
TanStack shipped a 1.0 release covering 76 packages with stable support for agents, subagents, coding workflows, sandboxes, provider adapters, and an experimental multi-environment harness. Builders get a portable open-source surface for chat UIs and tool-calling agents without vendor lock-in.
releaseharnesssdkportableShipped
- releaseGitHubAgentOS 2026.10.23 adds Channels settings and cleaner harness session UX↗
use-agent-os
The daily AgentOS build brings the Channels screen into the desktop app with live adapter status and guided pairing, plus fixes so finished coding-harness children no longer appear stuck as Running. ACP harness cards and session bars continue to improve visibility of delegated coding runs.
releaseagentosharnessdesktopShipped
- articleHacker NewsAgent Safehouse brings deny-first macOS sandboxing to local coding agents↗
eugene1g / atombender
The Show HN post and creator notes describe a pure sandbox-exec policy generator that needs no containers or VMs, with pre-built profiles and a web policy builder. Power users can lock agents to the minimum permissions that still allow auto-updates and keychain while keeping everything local.
articlemacossandboxlocalRead