Graph engineering ? Or we can say agents on steroids....

Graph engineering came to life this week. For a year the agent discourse has been loops vs graphs. Loops are easy to ship and impossible to audit. Graphs are auditable but nobody wants to hand-author a topology for "investigate this incident", because the shape is only discovered while working. So I built the missing piece: the graph is authored by a model at runtime, and a deterministic admission gate stands between proposing it and running it. The flow, from a real run in the demo video: You type one English question: grapharc go "why did checkout latency spike at 09:14 UTC?" A local qwen3:8b proposes a topology: triage, four parallel evidence pulls, a correlate join, hypothesize, verify, report The gate checks the proposal against the registry, the policy, the remaining budget, depth and acyclicity. All checks run on every proposal, so the model gets the complete list of objections, not just the first Only an admitted graph executes. You watch it live in the browser, every node amber while running, green with its own token bill when done A proposal names node kinds from an allowlist you wrote. It carries no code, no arguments that reach anything. Renaming a denied kind does not evade the policy. Rejections come back as structured codes with remedies, and the planner replans against them. Everything lands on one append-only JSONL trace. Replay, diff, metrics, cost attribution and the live view all read that same file, so the dashboard cannot disagree with the audit trail. MIT licensed, built on LangGraph, runs fully local on ollama or against OpenRouter/OpenAI/Claude. GitHub: https://github.com/CodeGraphContext/GraphARC PyPI: pip install grapharc