SandsDX
Back to Research
AI & Automation6 min read

How I Run SandsDX on a GTM Harness

How a one-person consultancy runs on AI agents under written rules: what the agents do, what stays human, and three mistakes the rules caught.

Page Sands ·

Fifteen years building B2B SaaS revenue systems: a Microsoft preferred vendor, Drift, Avalara, Blackbaud, ConnectWise. About →

A GTM harness is the software and written rules around AI agents that do go-to-market work: what they can read, which tools they can use, which actions wait for a human, and how their output is checked before anyone acts on it.

SandsDX is one operator and a set of AI agents. The agents do research, CRM upkeep, analysis and first drafts. I set the rules, approve anything that leaves the business, and do the client work that needs judgment. The software and rules that hold that together are a harness. HubSpot’s customer harness announcement made the term mainstream. This piece shows one in daily use.

What it is made of

The agents run in Claude Code. Around them sits an operating layer with four parts.

A written charter. Plain files that state the rules: ask before anything external or irreversible, never expose a credential, keep each client’s context separate, cite every claim. The agents read it at the start of every session.

Guards that enforce the charter. The most important rules are also code, because written rules alone get skipped under deadline pressure. A guard blocks any command that would print a credential or read a file the rules put off limits. Another blocks work that crosses from one client’s folder into another’s.

Skills and method files. Repeatable work lives in written procedures the agents follow, the same way a playbook works for a team. My pipeline audit method is public, as the open-source plugin New Logo Growth: six layers in order, a finding format, a scoring rubric, and a library of research sources checked by hand.

Memory and handoffs. Each project keeps notes on decisions and open items, and each session ends with a handoff file, so the next session starts where the last one stopped.

What stays human

I approve anything that reaches a client or the public, anything that spends money, and anything that is hard to undo. The agents draft emails and I send them. They prepare a deploy and I approve it. They propose a deletion and I confirm it.

Three catches this month

Three examples from this month show what the checks catch.

A tag that was not there. During a public audit, a scraping tool reported a “noindex” tag on a competitor comparison page, which would have been a strong finding. The method requires metadata claims to be confirmed against the raw page. The raw HTML had no such tag; the scraper had read it from an embedded video player. The finding was dropped before anyone saw it.

A key that would have billed the wrong account. A script that asks AI assistants buyer questions picked up an API key already present on the machine instead of the one I had chosen. A one-question test run caught it before the full run of 60 calls. The method now says to run one question first and to clear any keys not chosen for the run.

A total that did not add up. In a client report, a revenue total quoted in the summary did not match the sum of its parts. The verify step recomputes every total before delivery. Fixing it took two minutes.

None of these were model failures in the usual sense. They were the kind of mistakes a careful analyst makes too. The harness caught them because the checks are written down and run every time.

What it changes

The work that used to take a week of research now takes an afternoon of review. My time now goes to deciding what matters for each client and approving what ships.

If you are deciding whether your team is ready for agents, start where I did: write down the rules and definitions first. How to evaluate a customer harness lists the questions to answer before you buy one.

Frequently Asked Questions

What is a GTM harness?

A GTM harness is the software and written rules around AI agents that do go-to-market work. It defines what the agents can read, which tools they can use, which actions need a human's approval, and how their output is checked before anyone acts on it.

What should stay human in a GTM harness?

Anything that reaches a customer or a public audience, anything that spends money, and anything hard to undo. In my setup agents research, draft, analyze and propose. I approve every message, publish, purchase and deletion.

How do you stop AI agents from making things up in GTM work?

Write the evidence rule down and enforce it. Every finding carries a source and a date, numbers without a source are marked unknown, and a second pass re-opens each source before anything ships. In practice the second pass catches a few errors per project.

Put this to work on your GTM

Thirty minutes. Bring the number you need to move and I'll tell you what I'd check first in your system.

Schedule a Conversation

Not ready to talk? Run the AI Visibility Baseline on your company →