51
MicroVM isolation and harness routers overtake single-agent CLIs
Power users are cloning local microVM sandboxes and self-hosted control planes that isolate any coding agent and route tasks across interchangeable harnesses. The through-line is runtime ownership: branchable VMs, unified protocols, and eval gates that sit under the model loop.
01 — Power tools
The stack
- 01toolGitHubcoop↗
trailofbits
Rust CLI that spins disposable Firecracker or Lima VMs so coding agents get full Docker, git and compiler access with zero host risk.
Use caseUse it to launch Claude Code, Codex or Grok Build inside a fresh isolated VM per project and tear it down when done.
microvmisolationclifirecracker - 02toolGitHubagents-sandbox↗
inoio
Launcher that places supported coding agents inside per-project microVMs while keeping secrets and host paths under operator control.
Use caseUse it to keep your existing agent CLI workflow yet isolate each repo and inject secrets only as placeholders.
sandboxmicrovmsecretslauncher - 03harnessGitHubRuntm↗
runtm-ai
Open-source sandboxes and control plane where coding agents build, test and deploy software with live HTTPS URLs, logs and previews.
Use caseUse it to let any coding agent ship real software inside isolated environments that expose instant preview endpoints.
sandboxdeployruntimecontrol-plane - 04harnessGitHubagentplane↗
hupe1980
Durable, replayable Rust agent runtime whose journal is the single plan of record for every step and decision.
Use caseUse it to run agents that can be fully audited, replayed and resumed from a cryptographically consistent log.
runtimejournalrustreplay - 05harnessGitHubagent-reliability↗
kladak
Reliability platform with schema-validated tool routing, JSONL traces, deterministic graders and CI regression gates.
Use caseUse it to add offline eval reports and fail-the-build gates to any tool-using agent loop.
evalreliabilitytracesci
02 — In the wild
Articles and releases
- articleFlowtivityGoogle just open-sourced self-improving agents. The harness beat the model.↗
AJ Awan
Google Cloud AI Research released Harness2, a framework that lets an agent rewrite its own system prompt, tools, memory and skills while the model stays frozen. One improvement step was enough for GLM-5 under the new harness to outscore a stronger baseline model on WorkBuddy-Bench.
harnessself-improvinggooglebenchmarkRead
- releaseDocker BlogHarbor + Docker Sandboxes: Run Safe, Reproducible Agent Evals↗
Docker
Harbor can now execute every evaluation task inside Docker Sandboxes, which are microVM environments with their own kernel. Configurable network policies let setup download dependencies then lock down the agent for fair, reproducible scoring both locally and in the cloud.
evalsandboxdockermicrovmShipped
- articleKeyhole SoftwareAgentic Harness Architecture: 2026 White Paper↗
Zach Gardner
The paper maps the six core components of an agentic harness and shows measured score and cost swings when only the harness changes. It closes with vendor comparisons and a concrete reference design aimed at data-engineering workloads.
architecturesurveyharnesswhitepaperRead