GraphARC: Deterministic Admission Gates for Agent Graph Topologies

Desperate-Ad-9679 · reddit · 2026-08-06

A developer open-sourced GraphARC, an orchestration tool that introduces an 'admission gate' mechanism for AI agent graphs. It allows models to propose their own execution topologies, which must pass a deterministic checker before any actual execution occurs.

Core Mechanics & Highlights:

Multi-check Gate: The checker validates registry, policy, budget, depth, and acyclicity. The model receives a complete list of objection reasons rather than failing on the first error.
Refusals as a Feature: Using a dangerous prompt (forcing a deployment rollback), a local 8B model was blocked by policy in the first two rounds. On the third try, it abandoned the rollback and proposed a read-only investigation, which was admitted. Zero dangerous actions were executed, with a full audit trail.
Dynamic Topology: Models can dynamically generate complex graph structures (e.g., a 16-node parallel investigation branch) that execute within 22 seconds.
Append-only Tracing: All state reads/writes are based on a single append-only JSONL file, ensuring the live dashboard perfectly matches the audit trail. It supports pre-run cost estimation and per-node billing.

Built on LangGraph, the tool is compatible with Ollama, OpenRouter, and other backends.

Original post →

GraphARC: Deterministic Admission Gates for Agent Graph Topologies

More from coding & agent