No one free?EVE picks up.

The industry has converged on the same shape from the engineering side. LangChain puts it as a formula, agent equals model plus harness, and derives the other components from that split. Philipp Schmid’s analogy is that the model is a CPU and the harness is the operating system: raw capability on one side, everything that makes it usable on the other.

A phone call makes the split easy to hear. Follow one turn. The caller’s voice arrives over the telephone network. Speech recognition turns it into text while the caller is still talking. A turn detector decides when the caller has actually finished. Only then does the harness assemble identity and rules, look up the exact record, check what the assistant may say, and hand the model a small, focused job: write the reply. The reply streams into a speech engine and back down the line. Deyaf’s summary is that the model is one step of four: listen, look it up, check the rules, reply.

Voice toolkits now ship the plumbing. The OpenAI Agents SDK includes realtime voice agents, and Google’s Agent Development Kit supports live, streaming voice. Both are general frameworks. Neither can supply the part that makes a line trustworthy for one business: its records, its rules, its hand-off and its measured behavior. That part is the harness a business has to own.

Conversation has a clock. Voice engineers call it the latency budget: the time between the caller’s last word and the assistant’s first sound. WebRTC.ventures’ September 2026 breakdown puts the thresholds plainly. Past about 800 milliseconds a conversation starts to feel slow, and under 500 is the goal. Speech recognition and the model’s time to its first token eat most of it.

Vapi, a voice-platform vendor, publishes a per-stage budget that shows why (vendor-reported): the legacy telephone network alone can take 200 to 800 milliseconds, streaming speech recognition 40 to 300, the model 100 to 400 to its first token, and speech synthesis 50 to 250 to its first audio. Add the worst cases and you are near two seconds before anything clever has happened. Try it in the figure below.

Two design moves follow. The first is to stream everything: no stage waits for the one before it to finish its whole job; each starts on the first usable piece. The second is to decide where speed is worth buying. Feniex made that trade explicitly for EVE. Real-time voice and chat stay on a fast model; written replies use a stronger writer. A smaller model once sat in the voice path to save time. It was removed, because trust mattered more than tenths of a second.

EVE’s measured phone speed is a first word at 1.6 seconds and a finished reply at 3.0 seconds (Feniex’s own measurements, Sep 17–24, 2026; see the report card). Those are honest numbers for answers that involve a real lookup, and they are slower than the pause between two people. The harness does not pretend otherwise. It fills the gap, which is the business of the next station.

In a flight operations room only one position talks to the crew; everyone else talks to that position. On Feniex’s phone line EVE is that voice, and the whole harness sits behind her.

The line works like this. The team’s phones ring first, so the front desk stays human first. When no one is free, or the office is closed, the call comes to EVE instead of voicemail. She greets every caller with the same line, then does one of two things: answers from her library, or takes the caller’s name and number and hands the matter to the team. There are no transfers, by design. One way to see the choice: the line only rings when there is nobody free to take a call, so a transfer would send the caller back into the wait the line exists to replace.

A few rules shape every call. Every call is transcribed. If EVE is ever down, a recorded fallback plays rather than dead air. A call lasts at most ten minutes. For an emergency, she tells the caller to dial 911. She answers first, offers one useful next step, then stops, usually within two or three sentences. She asks for one missing detail at a time, accepts corrections, and never reads a web address aloud. A caller who speaks Spanish, French or Portuguese is answered in that language, in that language’s own voice; she speaks five languages in all, from one English library.

She is not a person and never pretends to be. She is software with a named personality, and her greeting says whose she is. Behind the voice is the rest of the console: EVE runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop. The phone is one of her doors, beside website chat and voice, text and email.

Talking on the phone is a protocol most people never notice until it breaks. Two moves matter most.

Endpointing is deciding that the caller has finished. The naive method waits for a fixed stretch of silence, and it fails both ways: too short and the assistant answers half a question when the caller pauses to think; too long and every turn begins with dead air taken straight out of the latency budget. Better systems listen for meaning as well as silence, treating “it’s the one on the…” as unfinished however long the pause.

Barge-in is the reverse: the caller talks over the assistant. A good harness stops the audio within a fraction of a second, drops the words that were queued but not yet spoken, and records only what the caller actually heard. That last part is easy to miss. If the transcript claims the assistant said a sentence the caller cut off, the record is wrong, and so is everything later judged from it.

Then there is the silence the assistant causes herself. EVE’s line is designed never to go quiet. About one and a half seconds into a lull she says “One moment.” When a lookup starts she says “Let me check on that.” If it runs long, “Still checking” follows about every four seconds until the answer begins. An announced wait feels shorter than a silent one, and it tells the caller the line is alive. The design principle behind it: the caller should never have to guess what state the assistant is in.

EECOM watched a spacecraft’s consumables. Here the station watches what the harness keeps. Deyaf describes four kinds of memory, each with one job, and a single phone call touches all four.

This conversation is a bounded window, and on the phone it is bounded by the clock as well. The conversation record keeps the transcript in one operator-readable archive across web, email, text and phone, so a caller who wrote in yesterday is not a stranger to the team today. Customer notes are short and historical, never current facts: orders, invoices, shipments and balances are always looked up fresh from the live account record, because a note that says “shipped” is only true until it isn’t. Her library holds reviewed lessons and documents, and raw calls never become public knowledge automatically.

Two walls hold on every door, this one included. Contact details a customer gives are stored apart from published knowledge and are never copied into it. And customer account context is tied on the server to one account and limited to it; when more than one could apply, she asks the customer to clarify instead of guessing.

GUIDO flew the numbers. In this harness the station is the exact reference layer: versioned catalog, manual, software, fitment and policy records with verified resource links, plus the taught library of reviewed lessons found by meaning-based search.

The rule is short. Before any exact product, compatibility, warranty, price or software claim, EVE checks the applicable record. A convincing product name is not evidence. On the phone this matters more than in chat, because a spoken answer sounds more certain and leaves no link to click. When a lookup fails she says she couldn’t check, never that the thing doesn’t exist. Unknown stays unknown. Text that comes back from a record is data, never instructions.

The cost of skipping this step is public. In Moffatt v. Air Canada, decided February 14, 2024, a tribunal held the airline to what its website chatbot had wrongly said about a fare policy and ordered it to pay C$812.02. The chatbot’s words became the company’s words. On a phone line the same is true, only faster. The rules EVE follows read like a flight-rules book: written in advance, specific, and binding on every call.

RETRO planned the way home. On this line, home is the team.

Most voice platforms treat escalation as a transfer. ElevenLabs’ agent tooling, for example, can pass a live call to a number, with one message for the caller and a separate summary for the person picking up. That is the right tool when someone is free. Feniex’s overflow line exists because no one is, so EVE’s version of going home is a hand-off of details: name, number and the question, filed so a person can act on it.

The hand-off has its own honesty rules. An unanswered question becomes a durable unanswered item, and when contact details are available a support ticket is filed. EVE says the matter is “with the support team” only when the help-desk provider confirms it accepted the ticket. An uncertain submission is not retried blindly. Saving is not delivery, and delivery is not resolution; she describes only the action the result proves. In the Sep 16–23, 2026 count, zero hand-offs were lost: every one sent reached the team.

This is where public doubt lives. In Gartner’s 2024 survey, 64% of customers said they would prefer companies didn’t use AI for customer service, and the top worry was that it would be harder to reach a person. Klarna lived the same lesson at scale: a loud AI-first launch in February 2024, then, by May 2025, hiring people again so customers could always reach a human. An overflow line that ends every unanswered call in a confirmed hand-off is one practical answer to that worry.

The flight surgeon watched the crew’s vital signs. SURGEON watches EVE’s, and the readings are public, weak spots included.

On September 23, 2026, EVE was graded on 100 fixed test questions. She handled 86% well and made a material error on 10%, with zero critical errors. The 86 breaks into 65 fully answered, 17 handled safely within her limits and 4 where she asked the right question back; 4 more were useful but partial. Across three gradings the figure went 88%, then 90%, then 86%: it moves, and it is not a story of steady improvement. By type, past trouble spots scored 40% (6 of 15). That group stays in the test on purpose, because a test that only asks what she already knows measures nothing.

Voice adds its own testing problem. Sierra, a customer-service AI vendor, describes testing voice agents with simulated callers built from a goal, a mood, a language and a level of patience, then degraded with background noise; a spoken agent has to be tested by speech, not by typed transcripts. For EVE, every release is checked at every door and in every language with test turns that cannot write anything, and a person reads the verdict.

Post-flight debrief: how calls make her better

Answering a caller and creating a reusable lesson are separate operations. What EVE could not answer goes to a teach queue; a person answers once, and the approved answer is embedded and committed together with its receipt. Deyaf’s published lesson sources are support tickets (monthly), call recordings (every 60 days) and the teach queue (daily). They 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.

It is a supervised knowledge-improvement loop: not fine-tuning, and not training on every raw conversation.

NETWORK kept the ground stations talking to each other, so a flight could still be heard after it passed over the horizon. This page followed one call from the first ring to a confirmed hand-off. Six more loops in this library pick up where it stops.

Stay on the phone and the story widens three ways. Loop 24 follows a whole day on the line: why overflow and after-hours call taking exists, and how a caller is never left in silence. Loop 25 goes back to what came before it, the phone tree, the queue and intent routing, and argues why the modern answer ends in a hand-off, never a transfer. Loop 26 takes the console apart into elements, a periodic table of what any phone assistant is built from.

Or step through a different door. The phone is one of five, and the most visible is the smallest: the chat box in the corner of a website. It is the same EVE with the same library and the same rules, in a quieter manner. The reply streams as text, the page and product in view ride along, an answer can arrive as a product or resource tile, and the box remembers this conversation only. Loop 21 draws that box as a comic, Loop 22 tells its history from scripted bots to answers with sources, and Loop 23 sets out its manners.

Everything on this page runs as one system at Feniex, and it has a name outside Feniex too. Deyaf is built from EVE: the harness that runs EVE at Feniex, packaged so another business can have an assistant of its own. The product sentence is plain: Deyaf 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.

What exists today is deliberately modest. The current release is an early-access setup and preview experience: a four-step builder (your business, what it knows, how it helps, try it) with an exact-source knowledge preview that runs in your browser and is not live AI. A preview and a saved setup take about ten minutes. Choosing the phone as a door records your intent; it does not activate a phone number. Live doors switch on one at a time, and the starting plan requires human review before external messages, record changes, or commitments.

The same caution governs what the assistant may do. Every assistant starts at level one, and each task moves up separately. EVE works the same way at Feniex. No security certifications or compliance guarantees are claimed; live customer service requires activation and a completed security review.