The Box in the Corner

Look at the bottom-right corner of almost any business website and you will find it: a round button, a little bubble, sometimes a face. Click it and a box slides up with a greeting. It is the smallest thing on the page, usually a few hundred pixels tall, and it is often the first place a customer actually talks to the company.

That box used to be simple. It was a live-chat window with a person on the other end, or a decision tree that offered four buttons and a link to the help centre. Today more and more of those boxes are wired to a language model that can write a fluent answer to almost anything. That fluency is exactly the problem. A model on its own will happily answer questions it has no business answering, in a confident, friendly voice, on the company's own website.

This issue is about what should sit between the box and the model. We call it an Agentic Harness: the system that decides what the assistant knows, what it is allowed to say, what travels with each message, what it remembers, and what happens when it doesn't know. The box is the part visitors see. The harness is the part that decides whether the box helps them or embarrasses you.

Customers are not starting from trust. In a Gartner survey published in July 2024, 64% of customers said they would prefer that companies didn't use AI for customer service, and their top worry was that it would get harder to reach a person. Qualtrics reported in October 2025 that nearly one in five consumers who had used AI for customer service saw no benefit from it, and about half feared that AI would stop them reaching a human. Zendesk's CX Trends 2026 report found that 74% of consumers get frustrated when they have to repeat information. Put those together and a visitor judges the box on three plain questions: does it answer from something real, does it remember what I just said, and does it let me reach a person when it can't help? Each of those is a harness question, not a model question.

Neither bot was broken in the way software usually breaks. Both did what a language model does when nothing stops it: produce the most plausible next sentence. A plausible sentence about a refund policy is still a sentence about a refund policy.

The fix is not a better sentence generator. It is a rule, enforced outside the model, that no policy, price, warranty or compatibility claim leaves the box until the harness has checked the record it comes from, and a path to a person when the record is silent. That is the whole difference between the two pages you just read.

Caption: the definition

“An agentic harness is everything around the AI model: what the assistant knows, what it remembers, where it meets customers, what it may do, and how your team teaches it.”

Deyaf's definition. Go deeper: how an agentic harness works.

Engineers have been converging on the same split for a while. LangChain's "The Anatomy of an Agent Harness" (March 2026) puts it as an equation: the agent is the model plus the harness. Anthropic's "Building Effective AI Agents" (December 2024) spends most of its pages on the parts around the model, from the tools it can call to the checks that confirm its work, and names customer support as a natural fit for agents. The model writes the sentences. Everything that decides which sentences are allowed, and what they are allowed to be based on, is the harness.

In a chat box that harness has to work in a very small space and a very short time. There is no room for a form, a menu or a page of terms. The visitor types one line, and within a few seconds something has to come back that is true, useful and honest about what it is. The rest of this issue takes the box apart to show where each of those three jobs is done.

Take a working chat box apart and you find the same handful of parts on almost every website. Most of them are familiar. Two of them do most of the trust work, and both are small enough to be cut when a designer runs out of room: the one-line disclosure that says you are talking to an AI, and the "talk to a person" path that actually leads somewhere. Pick a part below, or tap it in the drawing, to see what it is for.

Notice that only one part of the drawing, the stream, is the model's handiwork, and even there the model writes into a frame it did not build. The greeting and disclosure are fixed words. The tile comes from a record. The hand-off path is plumbing. The composer is a door with a rule on it. That is the harness showing through the glass: most of what makes a chat box trustworthy is decided before the model writes a single word. The same pattern holds when the box grows into a website voice, a phone line or an inbox. The doors change shape; the room behind them doesn't.

When a visitor presses send, the message never travels alone. A harnessed chat box packs a small bag for it: the page the visitor is on, the product in view if there is one, and the conversation so far. That is why "does this one work on induction?" can be answered at all. On its own, "this one" means nothing. With the product page riding along, the harness knows exactly which record to open.

The bag is deliberately small. Anthropic's engineering team describes a model's context as an attention budget and argues for the smallest set of high-signal information that gets the job done. Chroma's "Context Rot" study (July 2025) tested 18 models and found that every one of them got worse as the input grew longer. A box that dragged every earlier message into every reply would get slower, costlier and less accurate. So the conversation history is bounded: recent turns ride along, and older ones stay in the conversation record, where the team can read them, without crowding the next answer.

What never rides along matters just as much. Another visitor's conversation, obviously. The visitor's contact details copied into the shared library, which should never happen. And old notes dressed up as current facts: a note that says "asked about an order last month" is history, not an order status.

Zendesk's CX Trends 2026 report found that 81% of consumers expect a conversation to carry on without going backwards. A bounded window with a proper record behind it is how a box meets that expectation without pretending to remember everything forever. Try it: change the page, send a few messages, and watch what gets packed.

Slow the box down and a single answer turns out to be a short comic of its own. Deyaf puts the middle of it plainly: the model is one step of four. Listen, look it up, check the rules, reply. Around those four sit the page context that arrives with the question and the hand-off that waits in case the record is silent. Here is the whole sequence for one ordinary question, drawn the way it actually runs.

Every chat box eventually meets a question it can't answer from the record. That moment decides whether visitors trust it. The bad version is the one Cursor's users met: a confident guess. The lazy version is a dead end: "I'm sorry, I can't help with that." The honest version is a hand-off that actually lands somewhere.

In a well-built harness the hand-off is a chain of small, checkable steps. First, the unanswered question becomes a durable item of its own, so it can't evaporate when the tab closes. Then, if the visitor leaves contact details, a support ticket is filed. Only when the help desk confirms it has accepted that ticket does the box say the words "with the support team". Until then it says what is true: that it is passing the question on. If a submission comes back uncertain, it isn't blindly fired again.

That sounds fussy until you look at what the alternative promises. "Your ticket has been created" before anything was created is the same mistake as the invented refund policy, just aimed at a different fact. A harness describes only the action the result proves. Saving is not sending. Sending is not solving. Drafting is not sending either.

Two more rules hold the hand-off together. The box never pretends to be a person, not to smooth over a handover and not to sound warmer. And the box claims no authority it doesn't have: it can't approve a return, change an account or place an order just because a visitor asked nicely. In Deyaf's words, answering is not acting.

Visitors notice. The Qualtrics study found that half of consumers fear AI will block them from reaching a person. A visible, honest "talk to a person" path is the most direct answer to that fear, and it only works if the steps behind it are real. Deyaf draws the whole flow, including what happens when she doesn't know.

Everything in this issue so far has been drawn from a made-up kettle shop. Here is a real one. Eve is Feniex's assistant, and the chat box on Feniex's website is one of her doors. Eve runs on the agentic harness Deyaf packages: her knowledge, memory, doors, rules and learning loop.

Her chat box does what the anatomy page describes. Replies stream in as they are written. The page and product the visitor is looking at ride along with the question. Products and resources come back as structured tiles. The conversation history is bounded. There is a website voice as well, and beyond the website she answers on the phone, by text and by email, with the same identity, the same library and the same rules at every door.

Her manner is the "with a harness" page in practice. She answers first, gives one useful next step, then stops, usually in two or three sentences. She asks for one missing detail at a time. She checks the exact record before any product, compatibility, warranty, price or software claim, and says only what the result proves. She never pretends to be a person. When she doesn't know, the question becomes a saved item, a ticket is filed when there are contact details, and "with the support team" is said only once the ticket is confirmed.

She also gets better in a way that can be checked. An answer the team gives can be reviewed and taught as a lesson, but only approved lessons are published, and she is measured on a frozen set of test questions before and after. It is a supervised loop, not training on every chat: the loop that makes her better. More about her at Meet Eve, Feniex's assistant.

Her numbers, weak spot included

The same harness answers Feniex's overflow and after-hours phone line, where the team's phones ring first. In her first week on the line (since 17 September 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 calls were handed to the team with the caller's details taken. That is a hand-off, never a transfer.

As of September 24, 2026 · Feniex's internal Eve 3.0 report · machine-judged and provisional. See Eve's report card, weak spots included.