llmhut - AgentVM · The Runtime Your AI Agents Deserve

// architecture

Five modules.
One kernel.

AgentVM handles the OS-level concerns so framework developers can focus on reasoning, planning, and workflow design. Every module works standalone or together.

01 PROCESS

Process Manager

Spawn, pause, resume, and terminate agent processes. Isolated lifecycles, resource limits, crash recovery.

02 MEMORY

Memory Bus

Working, persistent, and shared memory tiers. Pluggable backends — SQLite, Redis, PostgreSQL, S3.

03 TOOLS

Tool Router

Central registry with automatic discovery, permission enforcement, rate limiting, and sandboxed execution.

04 MESSAGES

Message Broker

Pub/sub and direct channels for inter-agent communication. Typed messages, priority queues, dead-letter handling.

05 SCHEDULE

Scheduler

Parallel, sequential, conditional, and event-driven task execution with dependency resolution.

// the idea

LangChain is an app.
AgentVM is its OS.

Every agent framework reinvents the same infrastructure — process lifecycle, memory, tool routing, scheduling. AgentVM is the shared runtime layer that sits beneath all of them.

Framework-agnostic — works with any agent stack

Batteries included, not required — use just what you need

Observable by default — every op emits structured events

Crash recovery & checkpointing built in

TypeScript SDK with full type safety

Get AgentVM Contribute →

inter-agent messaging

// Create a shared channel const channel = new Channel('research-updates', { type: 'pubsub', schema: { topic: 'string', findings: 'string[]' } }); kernel.createChannel(channel); // Agents subscribe and publish const agent1 = await kernel.spawn('researcher-1'); const coord = await kernel.spawn('coordinator'); coord.subscribe('research-updates', (msg) => { console.log(msg.from, msg.data.findings.length); }); agent1.publish('research-updates', { topic: 'LLM benchmarks', findings: ['GPT-4o leads reasoning', '...'] });

// the problem

Every framework
reinvents the wheel.

Shallow process models, incompatible ecosystems, wasted effort. AgentVM gives every framework a shared, battle-tested foundation.

Without AgentVM

how every framework operates today

✗

Every framework builds its own process model

Shallow implementations, no crash recovery

✗

Memory is an afterthought

Chat buffers. That's it.

✗

Tools are framework-specific

No standard, no portability, no permissions

✗

No inter-agent communication standard

Agents can't talk across frameworks

✗

Debugging is guesswork

No structured events, no observability

With AgentVM

the shared runtime layer

✓

Shared, battle-tested process lifecycle

Spawn, pause, resume, kill — with crash recovery

✓

First-class memory bus

Working, persistent, shared tiers. Pluggable backends.

✓

Universal tool registry

Permissions, rate limiting, sandboxed execution

✓

Built-in message broker

Pub/sub + direct channels, typed messages

✓

Structured events for full observability

Plug in any monitoring tool and see everything

AgentVM Evolution Capability Unlock Status

v0.3.x — Ecosystem & Interoperability Framework adapters (LangChain, CrewAI), Python SDK, sandboxed execution, and resource monitoring. Moves AgentVM toward a universal runtime across AI ecosystems. Next

v0.2.x — Agent Execution Engine Fully functional orchestration layer with ToolRouter, MessageBroker, Scheduler, and pluggable memory. Introduced LLM-native agents, MCP integration, pipelines, and built-in tools for real-world execution. Active (v0.2.2)

v0.1.x — Runtime Foundation Core agent runtime with process lifecycle, in-memory state, event system, CLI, and TypeScript SDK. Established agents as structured, controllable processes. Shipped

v1.0 — Production Infrastructure Distributed execution, Kubernetes-native deployment, security hardening, and observability dashboard. Establishes AgentVM as production-grade AI agent infrastructure. Planned

LangChainCrewAIAutoGenOpenAIAnthropicGeminiGroqLlamaIndexVercel AI SDK

// also from llmhut