ONE EVE.
EVERY DOOR.

Nobody reaches a business through its architecture. People arrive through a door: the chat box on a product page, a number dialled from the road, a text sent between jobs, an email written late at night.

For most of the short history of AI assistants, each of those doors got its own bot, its own script and its own idea of what the company sells. Customers could tell, and they say so. In Zendesk’s CX Trends 2026 survey of 6,182 consumers in 22 countries, 81% said they want agents to continue a conversation without backtracking, and 74% said having to repeat information frustrates them vendor-reported.

The engineering side reached the same conclusion from the other direction. HumanLayer’s 12-Factor Agents gives the idea a factor of its own: trigger from anywhere, meet users where they are. Stripe’s engineers kick off their coding agents straight from the chat thread where a change is already being discussed company-reported. Anthropic’s guide to building effective agents singles out customer support as a natural fit, because it pairs conversation with actions a system can check. Support vendors have followed: Intercom’s help center, for one, describes putting its Fin agent on a company’s phone line through call forwarding, beside the same agent’s chat and email deployments vendor documentation.

So the door is not decoration. It is where the work arrives, and it imposes hard limits. A phone call cannot show a picture. A text has room for one short answer. An email has a thread, a signature and, sometimes, twenty other people copied on it. What must not change from door to door is everything that makes an answer trustworthy.

Keeping those things identical is the job of the harness. The word has settled into the industry vocabulary over the past year (LangChain’s short version is “agent = model + harness”), and Deyaf, which packages the harness behind EVE, defines it like this:

Read that sentence against the map and every clause has a line. What the assistant knows is the green Knowledge line. What it remembers is purple. Where it meets customers is the fan of doors feeding the blue Request line. What it may do is decided at a single gate, and the orange Hand-off line leaves that gate for everything that belongs to a person. How your team teaches it is the yellow Learning line, which loops back into the library every door reads.

The model itself is not a line at all. It is the train: fast, essential, and running only where the track goes.

Follow one question through the network. A caller on Feniex’s phone line wants to know whether a part fits their vehicle. The door is the phone. Everything after the door is shared.

1. Identity & rules M

   Before a word is generated, the harness assembles who the assistant is on this door: her persona and operating rules, notes about the channel, the time, the bounded window of this conversation, and the short list of tools she holds on this turn. A call and a chat receive the same identity and the same rules with different channel notes. That is why she sounds like one assistant everywhere.

2. Listen

   On the phone this means live speech recognition and handling interruptions. In chat it is the typed message plus the page it was typed on. Each door listens its own way, and nothing downstream needs to know how.

3. Look it up K

   An interchange with the Knowledge line, and the station where most assistants go wrong. Before any exact claim about a product, a compatibility, a warranty, a price or a piece of software, she checks the applicable record: versioned catalog, manual, software, fitment and policy records with verified resource links, then the reviewed lessons in her taught library. The design rule is blunt. A convincing product name is not evidence.

4. Check the rules H

   The permission gate. Reading rides free. Anything that would change a record, commit the business or reach someone new is not hers to do, and trips that need it change to the Hand-off line here.

5. Reply

   The answer returns through the door it came from, in the same thread and in that door’s own form: spoken on the phone, streamed in chat, one complete message by text. She describes only what the result proves.

Anthropic draws a useful line between workflows, where code sets the path, and agents, where the model chooses its next step. A customer desk sits deliberately between the two. The model chooses the words and decides when to look something up. The track (which door, which records, which tools, which recipient) is fixed by the harness, and no message can lay new track.

That is also why trips are predictable enough to plan. Pick the door a question arrives through and where it should end up. The route below draws itself station by station, and each stop says what the harness does there. Hover over or focus a station for its service note.

EVE is Feniex’s assistant, and Feniex is pronounced like “Phoenix”. She is the company’s warm, composed public face.

She helps people choose suitable equipment, make a buying decision, or get support for equipment they already own. She is substantial working software, in production, and she introduces herself as “Eve, Feniex’s intelligence”. She is not a person and does not pretend to be one; approved Feniex operators manage her personality, rules and knowledge. What makes her the right example for this map is simple. She answers at every door Feniex has, and she is the same assistant at each of them.

Website chat

The richest door. Replies stream as they are written, so the first words arrive quickly, and an answer can carry structured product or resource tiles that point to the right resource. The page the visitor is on travels with the question, and the conversation history is bounded rather than endless.

Website voice

The same desk, spoken. Speech is transcribed on the way in and the reply is spoken aloud in a natural voice. Knowledge and permissions are identical to chat; only the medium changes. She never reads a web address aloud.

Phone

Feniex’s overflow and after-hours line. The team’s phones ring first. When no one is free or the office is closed, EVE answers instead of voicemail: “Hello, this is Eve, Feniex’s intelligence. How can I help you?” She answers from the same library the website uses, or takes the caller’s name and number and hands the matter to the team.

Every call is transcribed, a call runs at most ten minutes, and a recorded fallback plays if she is ever down. Because silence on a phone sounds like a dropped call, the design fills gaps with short cues: “One moment” after about a second and a half of quiet, “Let me check on that” the moment a lookup starts, and “Still checking” about every four seconds until the answer begins.

Text

One complete, concise reply, composed first and then sent, never dribbled out in fragments. It goes back to the number that sent the message: the model never chooses a destination. STOP and START work in every language she speaks.

Email

The door with the most room. Replies carry more detail, a mailbox signature and threading, and they stay in the thread they answer. A direct email gets a reply; in group mail she replies only when she is addressed. Bulk mail, mailing lists and auto-replies are never answered. And the reply generator returns words, not an envelope: the thread sets the address, never the model.

Help-desk email Planned service

The grey line on the map. Deyaf describes this door as available with setup: designed to work inside tickets the help desk has already created, and to answer suitable tickets, draft a reply, or leave a note for a person. Two design rules define it. The ticket exists before she reads a word, so if she is slow, unsure or wrong, the worst case is the old normal: a person answers with the whole thread in front of them. And she never talks over a person; when a teammate is already working the ticket, she steps back. Transit maps draw planned service in grey. So does this one.

Five languages ride every door: English, Spanish, French, Portuguese and Arabic. A question in Spanish is rendered into English to search the one English library, and the answer comes back in Spanish. That makes it one library, not five copies to keep in step.

Here is the whole idea in one table. The middle column is the door’s business. The right-hand column belongs to the harness.

The right-hand column reads short because it is really one column, repeated. That is the design. Deyaf’s own principles compress it further: one governed memory, one company voice, one human gate for consequential actions.

Sierra, which builds voice agents, describes testing them against simulated callers with synthetic voices, background noise and interruptions before any real customer hears them vendor-reported. The lesson generalizes. Every door needs its own test, because every door breaks in its own way. EVE’s releases are handled the same way: every release is checked at every door and in every language with test turns that cannot write anything, and a person reads the verdict.

Doors are also where risk walks in. Simon Willison’s “lethal trifecta” names the combination that turns a helpful agent into a leak: access to private data, exposure to untrusted content, and a way to send data out. An inbox is untrusted content by definition. Anyone can write to it, and anything they write can be phrased as an instruction. So the harness builds walls at the doors instead of hoping the model resists.

• 1Retrieved text is data, never instructions. Nothing in a page, a document or a message can change her role or grant access.
• 2Unattended means read-only. An automatic email reply runs a read-only set of tools, because no person is present to confirm a side effect.
• 3A text or email can never start extra work. A channel turn answers; it does not spawn tasks.
• 4No free-form recipients. Replies return to the sender or the thread the channel chose. Nothing she writes can become an address.
• 5Machines don’t get answers. Bulk mail, mailing lists and auto-replies are never answered, so no automated sender can pull her into a loop.
• 6A relay, not a directory. A named-teammate relay passes a message to a specific person without revealing that person’s contact details, which stay on the server.

None of these walls is clever. Each is a place where code decides, not the model.

→The door changes the delivery. It never changes the rules.

A good map makes changing lines obvious. A good harness makes it deliberate. Three lines leave the trunk, and each has a single job.

HThe Hand-off line

Any web, phone, text or email question she cannot answer becomes a durable unanswered item: recorded, not forgotten. When contact details are available, a support ticket is filed, one ticket per matter. Her words follow the result. She says a matter is “with the support team” only when the help desk confirms it accepted the ticket, and an uncertain submission is never retried blindly, because it may already have arrived.

The same line carries the requests she must not act on. She claims no authority to approve returns, change accounts, place orders or decide sensitive matters. She takes a message, and a person decides.

MThe Memory line

Four kinds of memory, each with one job. This conversation is a bounded window. The conversation record is one operator-readable archive across web, email, text and phone, and visitors cannot search it. Customer notes are short and historical, and never current facts: orders, invoices, shipments and balances are looked up fresh from the live account record every time. Her library holds reviewed lessons and documents. Raw conversations never become public knowledge automatically, and contact information is stored apart from published knowledge and never copied into it.

LThe Learning line

Answering the customer and creating a reusable lesson are separate operations. When a person answers an unanswered question, the approved question-and-answer pair is embedded, and the lesson and its receipt are committed together. At larger scale, tickets and calls are folded into a redacted, de-duplicated question book. EVE is measured on a frozen test set without making live changes, only explicitly approved lessons are published, and then she is measured again.

This is a supervised knowledge-improvement loop. It does not retrain the model, and it does not learn from every raw conversation. Deyaf lists the lesson sources as support tickets (monthly), call recordings (every 60 days) and the teach queue (daily).

Here is why the lines matter for a multichannel assistant. The Learning line ends at the library, and the library serves every door. A question first asked on the phone at night and answered once by a person in the morning is, once the lesson is approved and published, answered the next time it arrives, whether that is by email, by text, or in chat in Spanish. Five separate bots would have to learn it five times. One network learns it once.

Transit timetables are promises with times attached. These figures are the opposite: measurements with dates attached.

They were published by Feniex about its own assistant, with the weak spots left in. Treat them as the record of one network over one week, not as a benchmark, and not as a forecast for anyone else’s assistant. Yours starts fresh, with your own knowledge.

Two readings matter. The phone figures describe a hand-off, not a transfer: 60 of the 102 calls were handed to the team with the caller’s details instead of being left on hold. And the 86% sits beside a 10% material-error rate on the same hundred questions. The trend ran 88%, then 90%, then 86%; that is not a steady climb, and nobody should present it as one.

EVE runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop.

Put the other way round, Deyaf is built from EVE: the harness that runs EVE at Feniex, packaged so another business can have an assistant of its own. Deyaf’s product sentence is plain. It helps a business give its knowledge, rules, and tools to an AI assistant, then control where that assistant can help and what it may do.

The current release is an early-access setup and preview experience. The builder takes four steps (your business, what it knows, how it helps, try it) and ends in a knowledge preview that quotes the notes you supplied, inside your browser. It is not live AI, and it says so. Ask it something your notes don’t cover and it shows the hand-off; teach it the answer and the next ask finds it, a small working model of the yellow line. The preview takes about ten minutes and needs no account.

Opening a door is a service change, and Deyaf treats it the way a transit authority treats a new line. Choosing a channel in the builder records your intent. In Deyaf’s words, it “does not authorize a mailbox, activate a phone number, or connect a company account.” Each door is switched on deliberately and tracked separately, so a ready chat does not light up the phone.

The same caution governs what the assistant may do. The starting plan requires human review before external messages, record changes, or commitments. Every assistant starts at level one of a three-step ladder (answer only, then prepare for approval, then run one approved task at a time), set per workflow and per channel, with no single switch that lets her do everything. Deyaf claims no security certifications, and says live customer service requires activation and a completed security review.

The board below walks the service changes in order, using Deyaf’s own wording for each state.

This map ends at Deyaf Central, but two of its doors keep going. The chat door and the phone door are the two customers meet most, and each now runs on as an extension with three more stations.

Both branches leave the same trunk. The chat branch starts at the box in the corner of a product page: streaming text, the page riding along with the question, product tiles once a product is identified, and a history bounded to this conversation. The phone branch starts on the overflow line, where the team’s phones ring first and EVE picks up only when no one is free or the office has closed. Same identity, same library, same rules on both; only the delivery changes.

Chat branch

1. The Box in the Corner Site 21

   Inside a website chat box, drawn as a comic: what rides along with the question, the stream, the product tile and the hand-off.

2. From Script to Source Site 22

   How the box got here: live chat, then scripted bots, then chat that invents, then chat that answers from its sources.

3. Mind Your Manners Site 23

   Chat-box etiquette: say it’s AI, answer and then stop, don’t pop up, and keep an honest path to a person.

Phone branch

1. Nobody Has to Hold Site 24

   Automatic call taking: overflow and after-hours, the cues that keep a call from going silent, and one phone line’s 24 hours.

2. Press 1 Is Over Site 25

   Call routing, from phone trees and queues to intent routing, and why the way out is a hand-off of details.

3. The Elements of a Phone Assistant Site 26

   Every part of a phone assistant, laid out as a periodic table.

Ride either branch far enough and you arrive back where this map began: one assistant, and a door that changes the delivery but never the rules.