# How to implement an agent system for mid-market

Helixfield’s production path for mid-market companies: name the loop, run a Field Survey, choose the current best platform door and method, seed institutional memory, wire only the tools the loop already uses (Microsoft 365, Salesforce, NetSuite, ERP when that is the work), evaluate on live work, teach the owner, and launch only after a human accepts the handoff.

1. **Name the loop.** Write the repeating work in one sentence and name the human who can reject the output. If you cannot name the loop, do not buy seats or start an install.
2. **Run a Field Survey.** Spend two to three weeks mapping the loop, the tools you already pay for, and the facts an agent would need. Helixfield prices this diagnostic at $9,500. It does not include a production agent.
3. **Choose the current door and method.** Pick the platform that fits access, data residency, cost, and owner preference — Grok bots, OpenAI Astra, Claude, Gemini, Cursor, a mixed bench, or what just shipped. The method we install now is Loop & Graph. When a stronger method lands, we move.
4. **Seed institutional memory.** Write the facts, rules, systems, people, and open questions the agent must pack. Label each proposed, accepted, or ask. Chat history is not the memory.
5. **Wire only the tools the loop needs.** Connect the applications the job actually touches. Write permissions down. An agent that can click every app is a liability.
6. **Evaluate on live work.** Score the agent on real jobs, not a scripted demo. Missing facts become asks. Invented numbers fail the evaluation.
7. **Teach the owner.** The owner learns the verbs of the current method: pack, act, extract, ingest. They must be able to reject a bad fact and refuse an outbound draft.
8. **Accept the handoff.** A named human accepts the system before it touches clients. Client-facing output stays drafted until someone says send. Thirty days of correction follow launch.