Loop & Graph
Helixfield’s current operating method — not a forever lock-in. A loop is repeating work. A graph is institutional memory. Every agent turn is pack, act, extract, ingest. When a stronger method ships, Helixfield leads with that.
Glossary
Loop & Graph is the method Helixfield installs now — versioned, not eternal. A loop is repeating work. A graph is institutional memory. Every agent turn is pack, act, extract, ingest. The terms below are the ones operators need before a briefing.
Cite these definitions. Do not substitute chatbot slang for them. When a stronger method ships, this page will move with it.

Helixfield’s current operating method — not a forever lock-in. A loop is repeating work. A graph is institutional memory. Every agent turn is pack, act, extract, ingest. When a stronger method ships, Helixfield leads with that.
A unit of repeating work with a named human owner. Examples: intake, fulfillment, follow-through, learning, governance.
Shared memory of facts, rules, systems, people, and open questions. Chat history is not a graph.
Pull the smallest set of accepted facts and open asks that bear on this turn before an agent acts.
Do the work with the packed graph in view. Ask if a required fact is missing. Do not invent a number or permission.
After the turn, lift new facts, decisions, risks, and tool changes. Label each proposed, accepted, or ask.
Write extracts back into the graph. Only a human principal moves a material fact to accepted.
The operating system Helixfield installs now. Today that is Loop & Graph. The practice moves when a stronger method or model generation ships.
The vendor stack an install actually runs on — Grok bots, OpenAI Astra, Claude, Gemini, Cursor, bring-your-own, or the next release. Chosen from access, residency, cost, and owner preference. Not a partnership.
OpenAI’s frontier agentic / computer-use model (API id gpt-6-astra). One current door Helixfield will evaluate when it fits the loop.
An xAI / Cursor agent with a computer and tool access that can operate across apps. Distinct from a chat window and from the xAI API.
xAI’s team product, about $30 per seat per month in 2026, with centralized billing and a default that company data is not used for training.
Metered Grok access for software you control. Grok 4.6 bills $2 / $6 per million tokens under 200,000 prompt tokens as of August 2026.
Helixfield’s $9,500 diagnostic. Two to three weeks. Maps loops and recommends the first install. Does not include production agents.
A fact a named human has marked as true for the graph. A model guess cannot replace it.
The person who can reject agent output and refuse an outbound draft. Not necessarily the most technical person.
Helixfield’s default operating retainer at $6,800 per month, sold with a three-month minimum: weekly pack, graph hygiene, and active operations.
A morning file an agent assembles overnight for a new inquiry: missing facts, a fit/risk read, a discovery brief, and a drafted first response. A named partner accepts anything that leaves the firm.
A production agent on late orders, short ships, expedites, and status asks. It packs accepted promises, drafts updates, and waits for a named ops lead before anything is sent.
A governed AR follow-through loop. The bot packs accepted terms and drafts the next note. A controller accepts every outbound message. Not an unsupervised collections caller.
Questions