Agentic HarnessFrom Script to Source

Read me first

Open almost any business website and there is a small box waiting in the corner. It has been there, in one form or another, for most of the commercial web’s life, and it has been rebuilt from the inside at least three times. First a person sat behind it. Then a script. Then a language model that could talk about anything, including things that were not true.

This page is a short history of that box, told the way the box itself might tell it: as a desktop you can poke around. Each era gets a window, and each window runs the same customer question, so you can watch what changed and what broke. The history is deliberately hedged. Dates appear only where a public source pins them down, and every conversation you see is an illustrative example written for this page, not a transcript.

The fourth window is the one this library is about. An Agentic Harness chat box looks like the other three from the outside: the same box, the same cursor, the same text arriving word by word. The difference is behind the glass. It answers from sources it can check, it knows which page you are on, it keeps the conversation bounded, and when it doesn’t know, it says so and hands the question to a person. Getting there meant keeping what each earlier era did right and fixing what each one did wrong.

The working example is EVE, Feniex’s assistant, who answers in the chat box on Feniex’s website and at every other door the company has. She arrives near the end. First, the question.

Windows.pic

Era 1 · live chat, staffed by people

At first, the box was a person.

The earliest chat boxes on business websites were a window into a support queue. A visitor typed; somewhere, a trained person read the message, checked what they needed to check, and wrote back. Live-chat software made this cheap enough to put on almost any site, and on plenty of sites it is still there today.

It got a surprising amount right. A person can read a messy question and work out what is actually being asked. A person knows which policy is current, can open the order, and can say the most useful words in customer service: “let me check.” When a question is unusual, a person notices that it is unusual and asks someone else. None of that is automatic. It is judgment, and judgment was the product.

What it got wrong was everything around the person. The box kept office hours, so late on a Sunday it turned grey and offered a form. It had a queue, so “you are number 6 in line” became the first thing many customers read. One agent often juggled several chats at once, which stretched every reply. Two agents could answer the same question two different ways, and a good answer, once given, lived in a single transcript where nobody could reuse it tomorrow.

What an Agentic Harness keeps from this era is the part people trust most: a real path to a person, and the honesty to use it. The harness does not pretend to be the agent who went home. It answers what it can prove, and when it can’t, it turns the question into something a person will actually pick up. That hand-off is not a failure state. It is the best idea the first era had, kept on purpose.

Era 3 · language-model chat with nothing around it

Then the box learned to talk, and to guess.

Large language models fixed the script’s biggest problem almost overnight. Suddenly the box could read any wording, in any tone, and reply in fluent, friendly sentences. Nobody had to write a branch for every question. For a while it looked like the end of the story.

It wasn’t, because fluency is not evidence. A model on its own produces the answer that sounds most likely, and in customer service the most likely-sounding answer is often a policy the business never wrote. Worse, a bare model will happily describe actions it has not taken: approved, refunded, shipped. With no record to check and no rule about what it may promise, the chat box became the most confident voice in the building and the least accountable one.

The public record from this era reads like a stack of error dialogs. Each case below is real and cited, and each one points at a specific piece of a harness that would have caught it.

• The $1 Tahoe

  Chevrolet of Watsonville · December 2023

  A user told the dealership’s chatbot to agree with anything the customer said. It then agreed to sell a 2024 Tahoe for one dollar and called the offer legally binding. The bot was pulled and the “deal” was not honored.

  Control: the model never makes a commitment the system has not approved, and text typed into the box is data, never orders.

  AI Incident Database #622

• The bot that swore

  DPD · January 2024

  After a system update, the parcel company’s chatbot swore at a customer who couldn’t get help and wrote a poem calling the company useless. DPD disabled the AI part of the bot.

  Control: tone rules, regression tests after every update, and an easy route to a person.

  ITV News, 19 Jan 2024

• The bereavement fare

  Moffatt v. Air Canada · decided February 14, 2024

  The airline’s website chatbot told a grieving customer he could claim a bereavement fare after travelling; the airline’s own policy page said otherwise. The tribunal rejected the idea that the chatbot was responsible for itself, held the airline responsible for the information on its website, and awarded C$812.02.

  Control: answer only from the current policy source, and send exceptions to a person.

  McCarthy Tétrault analysis

• The city’s bad advice

  NYC MyCity · reported March 29, 2024

  New York City’s business chatbot told employers and landlords things that would break city law, on topics such as tips and housing vouchers. A small disclaimer was the main safeguard.

  Control: grounding in authoritative sources, escalation on legal topics, and consistency tests that ask the same question many times.

  The Markup, 29 Mar 2024

• The policy that didn’t exist

  Cursor support bot “Sam” · April 2025

  Asked why users were being logged out, the support bot explained the bug by inventing a one-device-per-subscription policy. Users cancelled, and a co-founder corrected it in public.

  Control: never invent policy. “I don’t know; let me get a person” beats a confident guess, and the box says plainly that it is an AI.

  AI Incident Database #1039

Notice what the fixes have in common. None of them asks for a smarter model. They ask for a source to answer from, rules about what may be promised, tests that run after every change, and an honest way out. That is the work of the system around the model, and it is where the fourth era begins.

The working example

About this EVE

EVE is Feniex’s assistant, and the clearest way to see a fourth-era chat box is to look at one that is working. EVE runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop. You can meet EVE, Feniex’s assistant on Deyaf’s site.

On Feniex’s website, EVE’s chat box streams its answer as it is written, knows which page and product the visitor is looking at, shows product and resource tiles, and keeps a bounded conversation history. There is a website voice too. Behind every door (website chat, website voice, phone, text and email) it is the same EVE, with the same identity, the same library and the same rules.

Her manner is the history on this page turned into habits. She answers first, gives one useful next step, then stops; replies are usually two or three sentences. She asks for one missing detail at a time. Before any exact product, compatibility, warranty, price or software claim she checks the applicable record, and she says only what the result proves: saved is not sent, and sent is not resolved. She never pretends to be a person.

When she doesn’t know, the question doesn’t vanish. An unanswered question becomes a durable item for the team; when the visitor has left contact details, a support ticket is filed, and she says it is “with the support team” only once the ticket is confirmed accepted.

Read those numbers as a pair. On 100 fixed test questions graded September 23, 2026, 86% were handled well and 10% contained a material error; the second figure is the reason the first is not the whole story. In speed terms, that is about 3 seconds to a finished reply in website chat and on the phone, which Deyaf’s homepage walks through as four steps in about three seconds.

The same harness answers Feniex’s phone when the team is busy or the office is closed. In her first week on the line, since September 17, 2026, EVE answered 102 calls: 86 while the team was busy and 16 after the office closed. 93% of callers stayed and talked, and 60 were handed to the team with the caller’s details taken; that is a hand-off, never a transfer, because EVE does not transfer calls by design (Feniex’s own measurements, Sep 17–24, 2026; see the dated report card). A chat box and a phone line look like different products. Under a harness, they are two doors into one assistant.