DEV Community

Cover image for GO-GATE: Database-Grade Safety for AI Agents
William Louis Park
William Louis Park

Posted on

GO-GATE: Database-Grade Safety for AI Agents

#security #programming #ai #devops

Why autonomous AI systems need Two-Phase Commit (2PC) guarantees

The Problem: AI Agents Without Safety Rails

After watching agentic workflows rack up runaway cloud bills and attempt unsafe operations, I realized most existing frameworks (AutoGPT, BabyAGI, etc.) provide action capability but lack safety brakes.

AI agents that act without control are dangerous.

AI agents that wait for approval for everything are useless.

I built GO-GATE to solve this.

What is GO-GATE?

GO-GATE is a security kernel for AI agents that brings database-style Two-Phase Commit (2PC) guarantees to agent operations:


PREPARE → PENDING → COMMIT / ABORT
Enter fullscreen mode Exit fullscreen mode


`

Core Features

Feature Description
Two-Phase Commit Ensures dangerous operations don’t execute accidentally
Risk Tiers LOW (auto) / MEDIUM (verify) / HIGH (human required)
Fail-Closed Unknown operations default to human approval
Sandboxed Execution No shell injection, path traversal prevention
Immutable Audit SQLite WAL, append-only logging

Quick Start

bash
pip install go-gate
`

python

import asyncio
from go_gate import GoGate

async def main():
gate = GoGate()

# LOW risk: auto-approved
result = await gate.execute({
    "op_type": "FILE_WRITE",
    "target": "./data/output.txt",
    "payload": {"content": "Hello World"},
})
print(result.status)  # COMMITTED

# HIGH risk: requires human approval
result = await gate.execute({
    "op_type": "GIT_PUSH",
    "target": "origin",
    "payload": {"refspec": "HEAD"},
})
print(result.status)  # PENDING_HUMAN_APPROVAL
Enter fullscreen mode Exit fullscreen mode

asyncio.run(main())

Runs Locally (No Cloud Required)

  • No cloud dependency
  • No API key leakage risk
  • Data never leaves your machine
  • Apache 2.0 open source

Comparison: Why Existing Frameworks Fall Short

Framework Safety Model Issue
AutoGPT No built-in safety Fully autonomous, can execute dangerous ops
BabyAGI No built-in safety Same issue
GO-GATE Fail-closed Unknown = human approval, never blind

Key difference: GO-GATE isn’t “another agent framework” — it’s safety infrastructure that plugs into any existing system.

Use Cases

  • Autonomous code generation agents
  • Automated DevOps workflows
  • Enterprise AI systems (compliance requirements)
  • Any AI application needing deterministic safety

Built By

A mechanical engineer from Norway 🇳🇴, on my own hardware.

GitHub: https://github.com/billyxp74/go-gate

Would love feedback on the architecture and security model!

Tags (dev.to): #python #ai #opensource #security #devops #machinelearning

Top comments (0)