[Open Source] Cybertron — a governance-first architecture for agentic AI systems

[tairona_nice]

While building real-world multi-agent systems, we kept running into the same issues:

• Agents multiply faster than accountability

• Tools, prompts, and skills start driving behavior instead of design specs

• Systems “work”, but no one can clearly explain why an action happened

  

So instead of building yet another agent framework, we designed Cybertron —
a governance-first architecture for agentic AI systems.

Core ideas:

• Explicit command authorization (orders / “Tiger Tally”)
• Clear separation between orchestration and execution agents
• Human approval gates for BD / MD / PRD / SA / SD / E2E
• Mandatory memory for failures and reusable patterns
• Tool-agnostic adapters (Copilot, Claude, Cursor, CLI, etc.)
• Skills are adapters, not authority

This is not a theoretical paper and not a demo.
It comes from maintaining a large production system (tens of thousands of LOC),
where agent chaos becomes a real operational risk.

Cybertron is now open-sourced as a reusable governance framework.

Repo:
Cybertron-Agentic-AI-Framework

I’m especially interested in feedback from people dealing with:

• agent orchestration

• responsibility boundaries

• tool / prompt control

• long-running AI systems in production

  Todo:

• Hacker.md is a guideline designed to prevent potential vulnerabilities in programs. Any updates will be documented here.

• Consequence Model: The “consequence_model_whitepaper.md” is a document aimed at preventing AI from making critical errors, such as executing commands like “rm -rf *”. This document is not only applicable to coding but is also referenced within programs. Since AI currently lacks an understanding of what is dangerous or irreversible, and until we develop comprehensive world models or AGI, this serves as a safeguard to minimize AI’s potential for making purely foolish mistakes.

• Skill: Within the “agents\” directory, there’s a new “AgentsVx_skill” file that allows skills to be incorporated as tools for agents. It defines who can use them, how they should be used, and these rules are standardized and defined by the strategist, Zhuge Liang.

• A simplified path to avoid having small tasks go through all processes

In this framework, the most critical role is "you,"Opimus”. You are responsible for the overarching mission—communicating with your strategist and other autobots

At the beginning, you might only need a strategist, a coder/debugger, and a UI designer to kick things off. As for deciding who takes on these roles—whether as your strategist, programmer, or UI designer—that’s entirely up to personal preference.

[recursionrecursion1]

I think you might need more than just .md files to do what you’re aiming for. Who knows though … I hope it works!

[tairona_nice]

Sure, so it include some datatracks, rules and protocol betweens robots.

1. Data Tracks:

   • TRACK_ID

   • FROM, TO

   • DIRECTIVE (what to do, in imperative verbs: GENERATE_CODE, DIAGNOSE_ERROR, …)

   • STATUS (SUCCESS / FAILED / BLOCKED / REJECTED, etc.)

Every task is a Data Track with a unique TRACK_ID and a clear contract.
All communication goes through Prowl (hub-and-spoke),
so agents don’t randomly talk to each other and drift.

The OS is built around three types of documents:

1. Constitution – agents/ALLSPARK.md

   • Global rules: safety, sandbox boundaries, Energon (resource) discipline

   • Who is allowed to change architecture and when

   • What is explicitly forbidden (destructive ops, wild refactors, etc.)

2. Protocol – agents/data_tracks.md

   • Text-based schema for requests and responses

   • Defines:

     • [TRACK_ID], [FROM], [TO], [DIRECTIVE], [PRIORITY], …

     • Response fields: [STATUS], [ERROR_CODE], [OUTPUT], [NOTES]

3. Role Guides – agents/roles/*.md

   • One file per Autobot:

     • PROWL.md, IRONHIDE.md, RATCHET.md, BUMBLEBEE.md, WHEELJACK.md, TELETRAAN-1.md

   • Each file explains:

     • What this agent should do

     • What it must not do

     • How it should interpret Data Tracks

The simplest way to verify it is to fork it and ask your LLM (such as GPT, Opus, Gemini, etc.) to review the files AGENTS_v5.6_Transformers.md and ALLSPARK_v1.0.md.