Skip to main content

Command Palette

Search for a command to run...

Omnigent is a meta-harness. Liel is an event-sourced memory file.

Updated
3 min read

Databricks announced Omnigent on 2026-06-13.

The official blog describes Omnigent as a meta-harness that sits above Claude Code, Codex, Pi, and custom agents. Its goal is to support agent composition, policy control, and team collaboration.

References:

This article explains where liel fits in that world.

What Omnigent solves

Omnigent is not just about making one agent smarter.

It addresses a higher-level problem:

  • teams use multiple agents,

  • each agent has a different harness or UI,

  • context is split across sessions,

  • policies, approvals, cost, sandboxing, and collaboration need to span tools.

That makes Omnigent closer to a control plane or collaboration layer for agents.

Liel should not compete with Omnigent

Liel should not become another meta-harness.

Omnigent likely owns:

  • agent composition,

  • harness switching,

  • live collaboration,

  • web / mobile / native UI,

  • policy and sandboxing,

  • approvals,

  • managed server workflows.

Liel should own a different layer:

  • single-file memory artifact,

  • graph-shaped knowledge,

  • Actor provenance,

  • append-only Event history,

  • Source / evidence links,

  • diff / review / export,

  • local-first portability.

In short:

Omnigent = combine / control / collaborate with agents
Liel     = record how knowledge was formed

Why Liel core should stay small

Looking at Omnigent, it is tempting to add Session, ToolCall, Review, Approval, and Policy to Liel core.

That would be a mistake.

Those are tool or harness concepts.

If Liel adds every framework-specific object to core, it will grow around today's ecosystem instead of staying useful for the next one.

The core model should stay lower-level:

  • Actor

  • Event

  • Node

  • Edge

  • Source

Omnigent sessions and tool calls can be mapped by an adapter:

{
  "author": "actor:omnigent:claude-code",
  "operation": "tool_call",
  "target": "file:src/main.py",
  "payload": {
    "session_id": "...",
    "tool": "shell",
    "command": "pytest"
  }
}

Why this is useful under Omnigent

A meta-harness is great at running and coordinating agents.

But long-lived memory artifacts are a separate concern.

Putting Liel underneath Omnigent gives you:

  • a .liel file that survives a session,

  • provenance across agents and harnesses,

  • Git-reviewable memory,

  • diff / export / inspect workflows,

  • an artifact that can be opened outside Omnigent,

  • local-first portability without a managed server.

This is complementary, not competitive.

Conclusion

Omnigent points to a future where multi-agent work happens above individual agent tools.

Liel should not try to become that UI or harness.

Liel should be the event-sourced memory layer underneath it.

Omnigent: combine / control / collaborate with agents
Liel:     store how knowledge was formed

That positioning keeps Liel useful for Omnigent, Claude Code, Codex, OpenHands, LangGraph, and future agent frameworks.

Links:

Liel - Portable Graph Memory for Local AI Agents

Part 2 of 2

Building a local-first, single-file graph database for AI agent memory, MCP workflows, and portable knowledge graphs.

Start from the beginning

I Built a Single-File Graph Database for Local AI Memory

Portable graph memory for MCP-powered local agents