The harness as interface

For forty years, software has assumed one thing about you: that you will come to it. Learn the menus, find the settings, check the dashboard. Every app is a room you walk into and operate by hand.

Models that can reason and act through tools invert that assumption. You state an outcome — prepare the quote, book the jobs, get the lab ready — and a layer of software plans the steps, drives your existing systems, and comes back when something needs your judgment. We call that layer the harness, and the working definition fits in one line: an agent is a model plus its harness.

The model reasons. The harness is everything around it that makes the reasoning useful: the tools it may use, the memory it keeps, the rules it must obey, and the points where it must stop and ask you first.

The multi-pane dashboard was the best answer software had while you were the only operator. Every tool got its own, and every Monday began with opening six of them. The dashboard assumes the human is the integration layer — the thing that carries context from one system to the next.

When a system can operate the tools itself, the dashboard changes jobs. It stops being the home and becomes a view: something the system opens, or generates on the spot, when you want to inspect a number or audit a decision. The home is the session itself — your stated goals, the work in flight, the approvals waiting on you, and a plain record of what changed.

Most software forgets you between sessions. A harness keeps two kinds of state. Memory is what happened: which buyers wrote in, what was tried, what worked. Ontology is the structure underneath: your accounts, your services, your preferences, the rules of the house — never email after 8 p.m., this claim is off-limits, these zip codes are out of territory.

Both must stay visible and editable. A system that remembers things about your business that you cannot see or correct is not an asset; it is a liability with a login screen. Silent drift in what the system believes is a trust failure, so proposed changes to that structure are surfaced like any other work product.

This is where capability compounds. Each week the system needs less re-explaining: it resolves names correctly, respects preferences without being re-told, and spots repeated patterns worth turning into standing skills. A feature can be cloned. A year of accumulated, correctable structure about how your business decides cannot.

The whole architecture turns on one distinction: reversible work can flow, and irreversible acts wait for a human. Reading your CRM, drafting a reply, assembling a report — reversible. Sending the message, spending the money, deleting the record, moving the machine — gated.

Send, spend, move, and delete default to policy-checked, human-approved paths, with drafts and dry-runs as the standard starting state. Approval is not a checkbox but a dial: approve every send, approve only the first send to a new audience, approve any spend above a number you set.

Two logs matter, and they are not the same log. A trajectory log is the detailed trace used for debugging and learning. An audit log is the minimal, tamper-resistant record of every gated action. Blurring them either exposes private reasoning or starves compliance, so a serious system keeps both, separately.

The paper’s quietest claim may be its largest: the same loop runs digital and physical work. To the harness, a CRM and an email account are tools — and so are cameras, thermostats, locks, and factory equipment. Prepare the outreach for the conference leads and get the lab ready for the demo are the same kind of sentence: an outcome, some constraints, and the gates that apply.

We build only the software side of this today. But the pattern does not change when the tools do, and we expect this is how physical AI will meet its operators, too.

Strip the vocabulary and the thesis is the product we ship. Your website and social channels are where buyers show intent. The system answers from knowledge you approve, remembers what buyers ask and what makes them commit, and brings you the decisions — the publish, the send, the spend — instead of burying them in a console.

That is also why ad budget in our world sits behind a gate. Creative earns spend by producing real buyers first; the system proposes, you approve the exact number, and the cap holds. The harness pattern and the commercial pattern are the same pattern: outcomes stated by you, execution by the system, irreversibility reserved for you.

The paper this adapts is a position piece. It reports no benchmarks, no user studies, and no measured lift, and it says so on its first page. Its claim is architectural: this is how interfaces should work once software can act. Whether the claim is right is an empirical question, and the paper lists the open problems — how to evaluate this kind of experience, how to grow the memory structure without silent errors, how to add new tools safely — as invitations rather than afterthoughts.

We hold ourselves to the same standard on this site: the numbers we publish are ones we can show you.

Written by Tingsong Dai, founder of SentientWeb. This is a plain-language adaptation of the position paper “The Harness as Interface: A Position on UI/UX for the Agent Era” (August 2026), trimmed and re-voiced for operators rather than researchers. The original remains the version of record for citation.