32
Sandboxes, control planes, and systems harnesses that power users actually wire in
Today's picks lean into clone-and-run infra under the agent: microVM and K8s sandboxes, meta-harness scaffolds, Go/Rust TUIs, and production SDKs. These are the pieces that turn a model into a durable loop without the usual closed IDE tax.
01 — Power tools
The stack
- 01toolGitHubCrush↗
Charmbracelet
Go-native terminal coding agent with polished TUI, LSP intelligence, multi-provider switching mid-session, MCP extensibility, and session persistence.
Use caseUse it to run provider-agnostic agentic coding sessions directly in the terminal with LSP-backed edits and MCP tools.
clituigomcplsp - 02harnessGitHubAgent Sandbox↗
Kubernetes SIGs
Kubernetes CRD and controller for isolated, stateful, singleton workloads tailored to AI agent runtimes and RL, with warm pools and stable identity.
Use caseUse it to declare and manage long-lived agent sandboxes on any K8s cluster with pause/resume and template claims.
kubernetessandboxcrdinfra - 03repoGitHubCubeSandbox↗
Tencent Cloud
RustVMM/KVM microVM sandbox service with sub-60ms boot, hardware isolation, under 5MB overhead, and E2B SDK drop-in compatibility.
Use caseUse it to self-host high-density, hardware-isolated code execution backends for agent tool calls.
sandboxmicrovme2brust - 04toolGitHubAIO Sandbox↗
agent-infra
Single Docker container packing browser, shell, filesystem, MCP servers, VSCode Server, and Jupyter for unified agent environments.
Use caseUse it to give coding or computer-use agents one shared sandbox with VNC, CDP, and MCP endpoints.
sandboxdockermcpbrowser - 05harnessGitHubStrands Agents↗
strands-agents
Open-source Python and TypeScript SDK for building production agent loops with lifecycle controls, MCP, memory, multi-agent patterns, guardrails, and evals.
Use caseUse it to replace hand-rolled agent loops with a portable, model-agnostic harness you control end-to-end.
sdkpythontypescriptmcp - 06harnessGitHubmetaharness↗
ruvnet
Scaffold that generates a focused, branded agent harness with npx CLI, MCP server, scoped memory, skills, and witness-signed releases for multiple host runtimes.
Use caseUse it to spin up a repo-specific harness that drops into existing coding agents with governance and provenance.
meta-harnessscaffoldmcpcli - 07toolGitHubQwen Code↗
QwenLM / Alibaba
Open-source terminal coding agent with auto-memory, skills, sub-agents, teams, MCP, plan mode, and multi-protocol support optimized for Qwen coder models.
Use caseUse it to run headless or interactive agentic coding workflows with free OAuth tiers and IDE plugins.
clicoding-agentopen-source - 08toolGitHubKimi Code CLI↗
Moonshot AI
Single-binary terminal coding agent with fast TUI startup, skills, MCP, and ACP IDE integration, model-agnostic beyond Kimi defaults.
Use caseUse it to install a zero-Node coding agent that reads, edits, shells, and plans inside any repo.
clicoding-agentbinary - 09harnessGitHubMission Control↗
builderz-labs
Self-hosted control plane for dispatching tasks, reviewing runs, tracking spend, and operating multiple agent runtimes.
Use caseUse it to operate fleets of coding agents with observability, spend controls, and task queues.
control-planeopsobservability - 10harnessGitHubCoze Loop↗
coze-dev
Full-lifecycle agent optimization platform covering development, debugging, evaluation, and monitoring.
Use caseUse it to instrument and iterate on agent workflows with playgrounds, evals, and production monitoring.
evalobservabilityoptimization - 11toolGitHubAgenta↗
Agenta-AI
Open-source workspace for building, versioning, and collaborating on agents and automations with MCP and self-hosting.
Use caseUse it to prototype and share agent workflows as a team with prompt and skill management.
workspacemcpself-hosted - 12toolGitHubZeroClaw↗
zeroclaw-labs
Single Rust binary agent runtime with multi-channel support, provider plugins, supervised autonomy, and OS-level sandbox options.
Use caseUse it to run a low-overhead personal or team agent across Discord, Telegram, and CLI with policy gates.
rustruntimemulti-channel - 13toolGitHubNanoClaw↗
nanocoai
Lightweight container-isolated AI assistant inspired by larger claw-style agents, designed for readable code and secure defaults.
Use caseUse it to deploy per-agent containers with shared rooms while keeping the codebase small enough to audit.
containersecurityassistant - 14harnessGitHubAgent Security Harness↗
msaleme
Adversarial testing harness with hundreds of executable protocol-level tests for MCP, A2A, and governance controls.
Use caseUse it to CI-gate agent endpoints by actively probing for policy bypass and tool poisoning.
securityevalmcp - 15harnessGitHubAgentTeams↗
agentscope-ai
Collaborative multi-agent OS using Matrix rooms for transparent human-in-the-loop coordination across heterogeneous runtimes.
Use caseUse it to run mixed OpenClaw, Hermes, and other workers in shared rooms with file exchange and audit.
multi-agentmatrixorchestration - 16repoGitHubHeadroom↗
community (via best-of lists)
Content-aware compression layer for tool outputs, logs, and RAG chunks that claims large token reductions for coding agents.
Use caseUse it to shrink context before it hits the model while preserving answer quality on long agent runs.
contextcompressiontokens - 17toolGitHubLiteLLM↗
BerriAI
Unified interface and proxy for 100-plus LLMs with routing, caching, budgets, and agent SDK gateways.
Use caseUse it as the model router and spend control plane in front of any custom harness.
proxyroutingmulti-provider - 18toolGitHubRoo Code↗
RooCodeInc
VS Code and Cursor extension in the Cline lineage with approval-gated modes, strong MCP support, and custom agent modes.
Use caseUse it for plan-then-act coding inside the editor with per-step gates and MCP tools.
idemcpextension - 19toolGitHubTrae Agent↗
ByteDance
Research-friendly Python CLI agent for software engineering with modular architecture and multi-LLM support.
Use caseUse it to prototype SWE-bench style loops or modular coding agents without heavy product shells.
cliresearchpython - 20harnessGitHubChorus↗
community
CLI harness pattern for multi-LLM peer review before shipping code, listed among rising terminal agent tooling.
Use caseUse it to add a multi-model review gate in front of any coding agent commit flow.
reviewmulti-llmcli
02 — In the wild
Articles and releases
- releaseGitHub / hermes-agent docsHermes Agent v0.21 Pantheon release adds Bot Mode and live subagent steering↗
Nous Research
The August 31 Pantheon release bundles Bot Mode into the desktop app, gives cron jobs memory continuity, enables mid-flight subagent steering, and expands the MCP command surface. It continues the self-improving skill loop while remaining fully self-hostable.
releaseself-improvingmulti-agentShipped
- articleAgitech / zero8 / harnesses.sh style writeupsThe harness, not the model: why power users keep rebuilding the loop↗
various engineering blogs
Recent builder essays argue that sandbox fabric, permission gates, context eviction, and eval loops now dominate reliability more than checkpoint swaps. They map practical scorecards for tool harnesses versus closed IDE shells and show why mid-tier open runtimes keep winning power-user workflows.
essayharness-engineeringanalysisRead