Composable harness kernels and parallel sandboxes power users actually own

VoltAgent

Open-source TypeScript agent engineering platform with core runtime, workflows, supervisors, MCP tools, memory adapters, guardrails, and a paired observability console.

Use caseUse it to spin up typed multi-agent teams with durable memory and production evals from a single create-voltagent-app scaffold.

Google

Code-first Agent Development Kit with graph workflow runtime, task delegation, HITL tool confirmation, multi-agent hierarchies, and deploy paths to Cloud Run or Vertex.

Use caseUse it to define deterministic agent graphs and eval sets in pure Python then ship them to any model backend.

SWE-agent team

Sandboxed remote execution layer for AI agents that keeps the same API across local shells, Docker, Modal, Fargate and parallel sessions.

Use caseUse it to run hundreds of coding-agent trajectories in parallel without rewriting agent logic for each backend.

aiming-lab

Lightweight governance wrapper that turns any LLM client into a governed agent via a 6-to-14 step pipeline, YAML constitutions, risk patterns and audit logs.

Use caseUse it to wrap an existing OpenAI-compatible client in two lines and enforce tool permissions plus cost tracking.

harbor-framework

Evaluation and optimization framework that runs arbitrary coding agents against Terminal-Bench and other datasets across thousands of sandboxed environments.

Use caseUse it to benchmark Claude-style or custom agents head-to-head on Terminal-Bench 2.0 with concurrent rollouts.

HKUDS

Lightweight open agent infrastructure with tool loop, 43+ tools, skills, memory, multi-agent coordination, permissions and a personal ohmo gateway.

Use caseUse it to run a self-hosted personal coding agent over chat channels that forks branches and opens PRs.

agent-sandbox

Self-hosted E2B-compatible sandbox platform for untrusted LLM code, browser use, computer use and multi-tenant isolation on Kubernetes.

Use caseUse it to give every agent turn its own isolated container with shell, VNC and MCP lifecycle control.

codejunkie99

Rust-native agent runtime, SDK and CLI that ships one binary for laptop, CI, remote Linux sandboxes or edge with HTTP session environments.

Use caseUse it to build a coding agent that edits, tests and PRs from a single native binary without a JS runtime.

QuantaAlpha

Repo-scoped research harness that builds function and module graphs so agents explore only needed context with lower token use on coding benches.

Use caseUse it to drive graph-guided exploration of large monorepos before the agent writes patches.

UKGovernmentBEIS

Flexible framework for LLM evaluations with agents, scorers, datasets and sandboxed tool use aimed at rigorous agent benchmarking.

Use caseUse it to define custom agent eval suites with reproducible scorers and parallel task runners.

ByteDance

Long-horizon SuperAgent harness on LangGraph with sub-agents, sandboxes, memory, skills and message gateway for multi-hour research and coding tasks.

Use caseUse it to orchestrate lead-plus-subagent research and code workflows that survive minutes to hours.

DeepSeek AI

Everything-is-a-plugin agent harness on Cordis where model, tools, loop, sandbox, storage and UI are all swappable plugins under MIT.

Use caseUse it to compose a custom agent loop by mounting only the plugins you need then run the local web UI.

SWE-bench

Canonical software-engineering benchmark and harness for issue-to-patch agent evaluation across real GitHub repositories.

Use caseUse it to measure pass@1 of your coding harness on verified GitHub issues with official Docker runners.

GAIR-NLP

Open agent benchmark suite focused on realistic multi-step agency tasks beyond simple coding patches.

Use caseUse it to stress-test long-horizon planning and tool use outside SWE-bench style unit tests.

harbor-framework

Benchmark measuring agents on complicated real terminal workflows with Harbor as the official runner.

Use caseUse it to score shell-native coding agents on multi-step terminal tasks with concurrent Harbor jobs.

agentscope-ai

Production Java framework for distributed long-running agents with workspace sandboxes, multi-agent spawn, session recovery and control-plane dashboard.

Use caseUse it to deploy enterprise multi-agent systems with Redis-backed memory and Kubernetes sandboxes.

fairyshine

Headless workspace-first agent engine exposing HTTP APIs for sessions, runs, actions, sandboxes and SSE streaming for embedded products.

Use caseUse it as the backend runtime under your own chat UI and tenant model for internal agent platforms.

Cobus Greyling

Minimal reference implementation of six harness components (tools, memory, context, planning, verification, guardrails) with CLI and Gradio demo.

Use caseUse it to study and fork a complete small harness stack powered by Nemotron before building your own.

lox

Minimal composable Go library that implements the core tool-calling agent loop with hooks and thread state.

Use caseUse it to embed a tiny provider-agnostic agent loop inside Go services without pulling a full framework.

RyanAlberts

Weekly-rescored curated list of 100+ harnesses with MCP server, llms.txt and JSON so agents can recommend harnesses by capability.

Use caseUse it as a live catalog and MCP endpoint to let your own agent pick the right harness for a task.