Skip to content

research: OS-like architectures for LLM agent orchestration (AIOS, LLMOS, etc.) #86

Description

@clonable-eden

Summary

Survey of existing work that applies OS kernel concepts to LLM agent orchestration — the same direction cekernel is taking.

Key Projects and Papers

AIOS: LLM Agent Operating System (Rutgers University)

The closest parallel to cekernel. A full OS kernel for LLM agents, accepted at COLM 2025.

OS concepts implemented:

OS Concept AIOS Implementation
Scheduler Agent request dispatching with advanced scheduling strategies
Context switching LLM context snapshot & restoration (text-based and logits-based)
Memory manager Runtime data management
Storage manager Persistent storage
Access control Resource isolation between agents

Key result: Up to 2.1x faster execution for serving agents built by various frameworks.

"LLM as OS, Agents as Apps" Conceptual Framework

Systematizes the analogy: LLM = kernel, context window = memory, external storage = filesystem, tools = devices, prompts = user commands.

LlamaIndex llama-agents

Each agent runs as an independent microservice, orchestrated by an LLM-powered control plane with message queue communication.

Composable OS Kernel Architectures for Autonomous Intelligence

Further vision: NeuroSymbolic kernel architectures embedding logical reasoning, probabilistic inference, and neural computation directly into the OS substrate.

Other References

How cekernel Differs

Most frameworks (AIOS, llama-agents, LangGraph) are Python-based general-purpose frameworks with daemons, SDKs, and runtime dependencies.

cekernel takes a fundamentally different approach:

  • Shell scripts + UNIX primitives (FIFO, git worktree, jq, gh)
  • Daemon-less — no persistent process, no server
  • Minimal dependencies — only standard CLI tools
  • Claude Code native — built as a plugin, not a standalone framework

The same OS concepts, realized through UNIX philosophy: small tools, text streams, composition.

Relevance to Open Ideas

cekernel idea AIOS equivalent Notes
Session Memory (#74) Memory manager + Storage manager AIOS separates runtime (memory) and persistent (storage)
Signal (#75) Scheduler interrupts AIOS scheduler can preempt agent execution
Process states (#76) Agent lifecycle in scheduler AIOS tracks agent states for scheduling decisions
Priority (#77) Scheduling strategies AIOS supports multiple scheduling algorithms
Swap (#78) Context manager (snapshot/restore) AIOS has text-based and logits-based approaches
Cron (#79) N/A AIOS is reactive, not scheduled

AIOS's design decisions are useful reference material when implementing these ideas.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    researchRelated research, papers, and prior art

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions