Kubernetes CRDs and self-hosted KVM runtimes outpace vendor CLI sandboxes

Google / rakyll et al.

Declarative high-throughput orchestrator for agentic tasks that uses Workspace, Gateway, Task and Model primitives on top of Agent Substrate and Kubernetes.

Use caseUse it to apply YAML manifests that spin up sandboxed agent workloads with network allowlists, suspend/resume and ax ssh observability at cluster scale.

Flesymeb

Framework that wraps existing coding-agent harnesses into multi-day planning-coding-testing loops with continual self-improvement.

Use caseUse it to let an agent autonomously develop a full playable game or complex project across dozens of iteration cycles.

paolo-perrone

Educational five-layer agent harness (execution boundary, sandbox, memory, verification, context) that runs with a scripted model and no API key.

Use caseUse it to understand and copy the exact failure modes each harness layer prevents before wiring them into production.

tinyhumansai

Open-source agent harness with local-first memory, orchestration and workflows packaged as a privacy-focused desktop runtime.

Use caseUse it to run personal agents that keep all state on-device while still coordinating multi-step workflows.

Agent-Field

Open-source software factory harness aimed at open models, providing CLI agent fleets and coding workflows.

Use caseUse it to stand up a local software factory that drives open-weight coding agents through standardized pipelines.

caelis-labs

Local-first multi-agent coding harness that binds model providers and ACP agents to specialized roles inside a shared session.

Use caseUse it to coordinate specialized coding roles (planner, implementer, reviewer) without leaving your machine.

aeonfun

Autonomous agent framework that runs unattended on GitHub Actions with self-healing skills and drives multiple coding agents.

Use caseUse it to configure once and let agents ship PRs on a schedule with no human approval loops.

Oren Yomtov / Accomplish

Researchers disclosed Heapjack and Overpatch, two flaws that let OpenAI Codex escape its sandbox without approval prompts. The write-up argues that isolation must sit below the agent reasoning layer, not inside the same process the model controls.

Read