In-process WASM OSes and K8s sandboxes that keep agents off the host

WasmAgent

Embedded compliance layer with WASM kernels, MCP firewall, capability manifests, signed AEP rollouts, and trace-to-training export.

Use caseUse it to wrap tool calls with policy vetting, emit signed evidence records, and feed verified trajectories into DPO or PPO pipelines.

Mayflower

Bash-compatible WASM shell runtime with VFS, Pyodide Python, 88 utilities, and K8s-scalable dispatcher for dense agent sessions.

Use caseUse it to back Deep Agents or custom loops with an in-process or remote bash-plus-Python sandbox that cannot touch the host.

MichaelS1011

Capability-based Wasmtime sandbox for MCP tools and untrusted code with fuel, memory, time, and I/O caps plus signed execution records.

Use caseUse it to run agent-generated or MCP tool code under hard resource budgets and produce auditable signed results.

NarendraPatwardhan / OpytAI

WebAssembly VM giving agents a full Unix computer with shell, files, processes, snapshots, host tools, and capability boundaries.

Use caseUse it to replace bag-of-tools agents with a programmable guest computer that can pipeline shell work and call host-bound tools safely.

iii-experimental / iii-hq

Self-evolving agent OS of narrow Rust workers registering functions and triggers on a shared engine with wasm and microVM sandboxes.

Use caseUse it to compose domain workers for reasoning, memory, security, and execution that evolve under governed self-modification.

FeynmanZhou

Web console workbench for Deep Agents Code with pluggable sandboxes (local, Cube MicroVM, Tencent) and normalized event tracing.

Use caseUse it to drive and audit a headless coding agent through a three-pane UI with session history, artifacts, and sandbox swaps.

julianthome

Minimal Docker Compose sandbox that launches pi or opencode with Ollama, tmux, and git session branching for reviewable runs.

Use caseUse it to spin a polished local terminal agent environment that auto-wires host Ollama models and keeps changes on a review branch.

Harn project

Language and runtime for defining agent tasks, prompts, tools, and rules with enforced permissions and saved work history.

Use caseUse it to author reusable agent workflows as Harn programs that the runtime executes with sandbox boundaries and session logs.

rivet-dev

Rivet shipped agentOS v0.2.20 with runtime fixes for WASM pipe reads and native sidecar packaging. The release keeps the in-process V8 plus WASM model with optional external sandbox mounts for heavy workloads.

Shipped

GIGAZINE

Coverage of the HarnessTax study comparing seven models across three harnesses on SWE-bench Lite and Terminal-Bench. Success rates stayed close while cost varied up to five times, showing harness choice can dominate spend.

Read