Sandbox-native harnesses and embeddable runtimes that own the loop

SuperagenticAI

Self-organizing multi-agent swarm that invents specialist roles and runs them in isolated sandboxes against a git repo.

Use caseUse it to turn a natural-language goal into parallel scout, engineer, and tester agents with emergent roles.

fabledruns

Lightweight local-first Go agent harness as a single binary with tools, skills, sessions, memory, and permissions.

Use caseUse it to run specialized local coding agents with fast startup and no account or telemetry requirement.

Mem0

Drop-in multi-level memory layer that extracts facts and retrieves them for agents and apps across sessions.

Use caseUse it to bolt persistent user, session, and agent memory onto any coding or assistant harness.

Letta

Stateful agent runtime in the MemGPT lineage where agents edit their own long-term memory with tools.

Use caseUse it to run agents that manage hierarchical memory as first-class OS-style state rather than prompt stuffing.

skarnati20

Slim extensible coding harness written in Common Lisp for REPL-native loops, macros, and custom tools.

Use caseUse it to experiment with macro-defined agent loops inside SBCL instead of a heavy TUI product.

Charmbracelet

Crush adds Plan Mode activated with shift-tab for collaborative planning before execution, plus performance work and headless reasoning-effort flags. Power users get subscription-backed ChatGPT paths and tighter non-interactive orchestration without leaving the terminal harness.

Shipped

Sergey Malenkovich

The essay defines the harness as the loop, tool dispatch, memory, approvals, and sandbox around the model, arguing most real exploits land in that layer. It maps CLI, IDE, and assistant harness types and frames security as governance of scaffolding rather than model breakage.

Read