Sandboxes and control planes that put the harness under your keys

AetherLabsAI

Training-free multi-agent framework for recursive self-improvement that explores environments, verifies outcomes, and stores reusable memory.

Use caseUse it to let fixed-weight agents discover and retain environment knowledge through curriculum, actor, and verifier loops.

steel-dev

Open-source browser API and sandbox built for AI agents, with sessions, stealth, proxies, CDP, and self-host or cloud options.

Use caseUse it to give any agent reliable web automation without building browser infrastructure or fighting anti-bot measures.

LiteLLM-Labs

Unified control plane and API that registers, runs, schedules, and observes agents across multiple runtimes from one UI.

Use caseUse it to expose OpenCode, Hermes, Deep Agents, and managed runtimes behind a single gateway for your team.

superhq-ai

Sandboxed multi-agent orchestration platform that runs coding agents inside microVMs with an auth gateway that never exposes keys.

Use caseUse it to launch isolated Claude Code or Codex workspaces on Apple Silicon with host-side credential injection.

Design-Arena

Model-agnostic open-source harness that runs the same prompt against parallel sandboxed coding agents for preference collection.

Use caseUse it to benchmark real multi-file coding behavior across providers and build preference ground truth.

StanfordAgent

Open-source terminal daemon and orchestration layer that spawns and monitors many coding agents across machines from one interface.

Use caseUse it to drive native Claude Code, Codex, Pi, and other processes with optional remote relay without wrapping their behavior.

sinameraji

Local open-source sandbox manager that creates, pauses, and runs agent environments on laptop, VM, or bare metal with Docker, Apple VZ, or Firecracker.

Use caseUse it to spin multiple isolated coding-agent workspaces with ephemeral tokens instead of injecting real API keys.

Avnit Bambah

Google Cloud released a portable Developer Plugin and Agent Plugins specification that bundles skills and MCP servers for any coding harness. The piece argues the guardrail layer traveling across runtimes matters more than another closed harness.

Read

Superagentic AI

SuperQode 2.4.0 adds TypeSafe AI's Jev System One model as an opt-in typed decision service inside the existing permission and control loop. Coding models still generate, while Jev answers frozen questions about state for ALLOW, DENY, or ASK.

Shipped

Niklas Palm

The essay walks from chat assistants through tool-calling and code agents to managed harnesses, then builds a lightweight code agent in a managed runtime with tracing. It argues the minimum viable harness is a code agent plus sandbox tools plus observability.

Read