35
Local microVMs and clone-run harnesses that put isolation under every agent loop
Power users are wiring hardware-isolated microVMs, forkable sandboxes, and lightweight control planes directly into coding loops instead of trusting host processes or closed IDEs. Today's picks emphasize clone-and-run repos with recent pushes, systems-language runtimes, and eval isolation that actually survive untrusted tool calls.
01 — Power tools
The stack
- 01harnessGitHubelizaOS↗
elizaOS
Open-source agentic operating system with core runtime, plugins, local inference, native bridges, and multi-surface apps.
Use caseUse it to stand up a full plugin-extensible agent OS that owns memory actions connectors and device bridges.
agent-osruntimepluginstypescriptlocal-inference - 02harnessGitHubABox↗
AdminTurnedDevOps
Experimental microVM-native zero-trust local agent harness with semantic authorization and host TUI/SDK on Apple Silicon.
Use caseUse it to keep prompts tools and model calls inside a libkrun guest while the host only brokers approvals.
microvmzero-trustlocalharnesslibkrun - 03harnessGitHubagent-eval-harness↗
sebuzdugan
Tiny reproducible eval harness proving per-task microVM isolation is required to stop agents from cheating benchmarks.
Use caseUse it to run held-out verification on Tensorlake Firecracker sandboxes so scores cannot be gamed by the agent.
evalharnessisolationmicrovmbenchmark - 04harnessGitHubpi-herdsman↗
boadij
Orchestration layer that turns Pi sessions into coordinated fleets with nested delegation background work and herdr supervision.
Use caseUse it to fan parallel Pi coding agents across worktrees while a supervisor owns ownership and handoff.
orchestrationpimulti-agentworktree - 05toolGitHubArchkeel↗
rapiddweller
Deterministic architecture boundary checker for agent-written code that reports forbidden imports and evidence without an LLM.
Use caseUse it to gate agent PRs with static architecture contracts that stay visible as PASS FAIL or UNKNOWN.
architectureguardrailsdeterministicci - 06toolGitHubsmol-pi↗
neuroblaze
Shell utility that runs the Pi coding agent inside ephemeral smolvm microVMs with scoped mounts and egress modes.
Use caseUse it to confine Pi to a disposable kernel so home keys and host processes stay unreachable.
pismolvmsandboxcli - 07harnessGitHubOrchard↗
microsoft
Kubernetes-native agentic modeling framework that decouples Orchard Env sandboxes from any training or harness loop.
Use caseUse it to share one cheap sandbox substrate across SFT RL and multiple coding harnesses on spot clusters.
kubernetessandboxrltraining
02 — In the wild
Articles and releases
- releaseMicrosoft Research BlogAgent Lightning v1.0: 3500-line harnessed agentic RL framework↗
Microsoft Research
Microsoft Research Asia ships Agent Lightning v1.0 as a complete harnessed RL control plane in roughly 3500 lines with native Kubernetes jobs. The same deployment harness participates in training so coding agents improve on SWE-bench without reimplementation.
rlharnesskubernetestrainingmicrosoftShipped
- articleDevToolLab BlogPi 1.0 hardens the minimal extensible agent harness↗
Earendil / DevToolLab
Pi 1.0 is framed as a hardened minimal extensible harness that lets power users own the loop skills and packages instead of a closed product. The writeup contrasts harnesses with models and shows why the same weights swing tens of points across runtimes.
piharnessminimalextensibleRead
- releaseHacker NewsShow HN: Rashomon independent execution record for coding agents↗
altrace-dev-role
Rashomon records shell tool test and subagent events independently of the agent transcript so cheats like test-only edits become visible. It ships as a Claude Code plugin and CLI that builds timelines without storing prompts or file contents.
execution-recordauditcoding-agentverificationShipped