August 31, 2026

3 Minute Read

See how the latest AI and digital advocacy tools can transform your work.

THIS WEEK'S EDITION

What an AI Harness is,

and Why You Should Care

Most conversations about AI right now are about models and agents. Which model is smartest. Which agent your organization should be paying for. Both conversations skip the layer sitting between them, and that layer is where much of the actual work gets done. In today's short read, we explain what a harness is, why most of the AI products you keep hearing about are harnesses rather than models, and why you cannot understand agents until you understand this one word.

The Genius Who Cannot Get Into The Building

Think of the model as a brilliant new hire on day one. Encyclopedic knowledge, sharp judgment, useful on any issue you put in front of them. But also no badge, no laptop, no access to the client files, no idea which coalition partners are sensitive, and no authority to send anything to anyone. Enormous capability, zero reach.

The harness is everything that turns that person into a functioning member of your advocacy shop. The building access. The files. The tools they are allowed to touch. The sign-off required before anything goes to a client.

The model reasons. The harness acts.

The industry has settled on a formula worth memorizing: Model + Harness = Agent. On its own, a model produces text and nothing else. It cannot open your software stack, take an action for you, or remember yesterday. Attach a harness and it can do all three.

The Harnesses You've Already Heard About

Here is the part that surprises people. The AI products driving agentic workflows are mostly not models. They are harnesses.

🔱 Claude Code. The harness for software development. It runs in a terminal or inside a developer's tools, reads an entire codebase, writes and tests changes, and grinds through long jobs unattended.

🔱 Claude Cowork. Claude Code's sibling, built for people who do not write code. It works on your real files inside a sandbox, connects to systems like email and Google Drive, and runs bespoke, scheduled tasks.

🔱 Claude Design. A harness aimed at visual work. Leave-behinds, one-pagers, slides, prototypes. You describe what you need, then refine through conversation and direct edits.

🔱 Codex. OpenAI's answer to Claude Code, now spread across a terminal, a desktop app, the browser, and GitHub. OpenAI has been openly repositioning it from a coding tool into a general enterprise agent platform.

🔱 OpenClaw. The open-source outlier. It runs on your own machine and you talk to it through WhatsApp or iMessage the way you would text a colleague. It went from a side project to one of the most-starred repositories on GitHub, and OpenAI hired its creator in February.

🔱 NemoClaw. NVIDIA's answer to the obvious safety objections. It wraps OpenClaw in a sandbox governed by a written security policy your IT team can actually review, and it can run on your own hardware so nothing leaves the building. Guardrails are now their own product category.

Different models sit underneath these tools, and several will run on whatever model you point them at. That is the tell. What separates them is not the intelligence inside, it is the harness built around it.

What This Means For Public Affairs

Two things. First, the playing field for who has the competitive advantage in AI is rapidly shifting. Models are becoming a commodity, and the opposition will have the same ones you do. What they will not have is your harness: the files it can see, the style it follows, the sourcing rules it obeys, the approval gate it cannot skip. Model access is a subscription. A harness is an asset, and it does not walk out the door when someone switches vendors.

Second, and Washington has not caught up here: nearly every AI bill on the table regulates models. But most of what people actually fear from agents lives in the harness: too much access, too few approvals, an action nobody authorized. Legislators are inspecting the engine while the risk sits at the steering wheel.

Three Ways To Operationalize This Email

🔱 Ask which harness before you ask which model. In your next vendor conversation, ask what the tool can touch, what it remembers, and what it needs permission to do. Those answers tell you more than any benchmark.

🔱 Move one real workflow out of the chat window. Take something recurring, a weekly clip report or a coalition update, and run it somewhere with access to your actual files. You cannot evaluate agents from a chat box.

🔱 Write your outer harness down. House style, sourcing standards, approval steps, the do-not-touch list. Today that is tribal knowledge. Soon it will be the instruction set that makes your agents worth having.

🔱 If you're enjoying this content, please consider forwarding this email to a colleague or friend.

🔱 If you're not already a subscriber, please sign up here to stay up to date on the latest developments in political technology.

🔱 If you have suggestions for future newsletters or any other feedback, we'd love to hear from you.