Declarative sandboxes and control planes that keep agent loops alive

loopx-project

Local-first control plane with durable state kernel for long-horizon agents and teams, goals, evidence, heartbeats, and cross-runtime handoffs.

Use caseUse it to keep multi-day coding or research goals alive across sessions with claims, leases, reward memory, and recoverable peer ownership.

paperclipai

Open-source app and server for managing teams of AI agents at work with org charts, budgets, goals, adapters, and cost tracking.

Use caseUse it to hire heterogeneous agents as employees, assign company goals, and track handoffs and spend from one dashboard.

HKUDS

Hub and harnesses that wrap existing CLI software into agent-native interfaces with per-app agent-harness modules and plugins.

Use caseUse it to expose Blender, GIMP, Ollama, n8n or dozens of other tools as structured agent skills without rewriting the apps.

deepclause

Lightweight Node.js WASM-based Alpine Linux VM for agents with shell, networking, firewall and persistence for isolated command execution.

Use caseUse it to run agent shell commands inside a worker-thread Linux VM with DHCP, port forwards and snapshottable state.

various / Terminal-Bench related

Generalized harness for evaluating and improving agents at scale, paired with terminal benchmarks.

Use caseUse it to run large-scale agent rollouts against Terminal-Bench style tasks with consistent isolation and scoring.

snapshot-fork microVM projects

Snapshot-fork microVM sandboxes that give agent sessions clean isolated starting states with declarative lifecycle and parallel execution.

Use caseUse it to fork agent sandboxes from snapshots so each trial starts clean and can run in parallel without shared state.

freakynit

Low-friction sandbox that attaches per-tool permissions and instant clonefile repo copies so agents can destroy a writable tree then sync back.

Use caseUse it on M-series Macs to let hesitant teams get best-effort isolation without full VM setup friction.

Microsoft

Training-oriented harness with optimization loops over agent rollouts for policy improvement rather than one-shot scoring.

Use caseUse it to close the loop from agent trajectories back into better behavior via APO-style training.

castor-ai / claytonia et al.

Secure execution layer that turns every agent tool call into a kernel syscall with budgets, immutable journals, and suspend/resume.

Use caseUse it to bound agent power with structural budgets and pause/resume without replaying costly early steps.