The periodic table of theAgentic Harness
Everything an automatic call-taking and routing system needs, laid out as matter: 28 elements in five families. The model is one tile. The table is the harness.
Before you handle anything
When a business says it has put AI on the phones, it usually means it bought a model that can talk. That is one element. What a caller actually meets is a compound: something that hears, decides when they have finished speaking, stops when they interrupt, looks things up, refuses to guess, answers in their language, knows when the office is closed, and knows what to do with a question it cannot answer. Take away any one of those and the call changes character, sometimes in a way nobody notices until a customer does.
This page lays the pieces out the way chemists lay out matter: as a table. Twenty-eight elements in five families: listening, thinking, speaking, routing and safety. It is our editorial sketch of automatic call taking and routing, not an industry standard and not a product specification. Its argument fits in one line: the Agentic Harness is the table; the model is one element.
“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. Read how an Agentic Harness works, end to end.
Others draw the same boundary. LangChain's engineers write it as an equation, agent equals model plus harness, in The Anatomy of an Agent Harness. Philipp Schmid compares the model to a CPU and the harness to the operating system around it (The importance of Agent Harness in 2026). On a phone line the operating system matters more than usual. A caller cannot scroll back, cannot see a typing indicator, and hangs up on silence. Every missing element is audible.
How to read the tiles: each carries a number, a symbol, a name and one key figure or moment. A filled dot marks an element that EVE, Feniex's assistant, is documented to carry on Feniex's overflow and after-hours line. An open ring marks general industry practice, described from published sources, with no claim that EVE works that way. One tile carries an empty set: an element EVE leaves out on purpose. Open any tile for its data sheet, or drop two or three into the reaction chamber further down and see what they make together.
Chat has a send button. A phone call does not. The first job of a phone assistant is to turn sound into words, and then to decide, without being told, that the caller has finished a thought. Voice engineers call that decision endpointing, and it is harder than it sounds. People pause mid-sentence to find a part number, trail off, or say “um” and keep going. Wait too long and the line feels dead; jump in too early and the assistant answers half a question.
Voice platforms publish how they manage it. Vapi describes starting the model when it guesses the caller is done and cancelling if they keep talking; Retell exposes interruption sensitivity as a setting (both vendor-reported). The mirror image of endpointing is barge-in: when the caller talks over the assistant, the assistant stops and listens. An assistant that cannot be interrupted feels like a recorded message with opinions. These are general industry practices. This page does not claim how EVE handles either one, which is why both tiles carry an open ring.
Two listening elements are about honesty more than timing. Repair is what happens when the words did not come through: ask again rather than guess. deyaf.com shows that line in all five of EVE's languages, starting with “I didn't catch that.” Clarify is the discipline of asking for one missing detail at a time; EVE asks for one detail at a time and accepts corrections. Language detection closes the family. EVE works in English, Spanish, French, Portuguese and Arabic, and callers in Spanish, French or Portuguese are answered in that language's own voice.
This is where most explainers stop, and where the table insists you keep going. The model, element 18, composes the reply. deyaf.com describes each of EVE's answers in four steps: listen, look it up, check the rules, reply. The model is one step of four. Deyaf's notes add that real-time voice and chat stay on a fast model while written replies use a stronger writer, and that a smaller model was once removed from the voice path because trust mattered more than tenths of a second.
Around the model sit the elements that make a reply worth trusting. Rules and identity come first: who the assistant is, what it may say, and that it never pretends to be a person. Memory is bounded: this conversation, plus short notes that are never treated as current facts. Orders, invoices, shipments and balances are looked up fresh from the live record every time. The library is taught: reviewed lessons, found by meaning, and raw conversations never promoted to public knowledge automatically. Lookup reaches the exact reference layer: versioned catalog, manual, software, fitment and policy records.
Then comes the element the others depend on: evidence. EVE checks the applicable record before any exact product, compatibility, warranty, price or software claim, and says only what the result proves. In Deyaf's words, a convincing product name is not evidence. When a lookup fails, the honest sentence is “I couldn't check”, not “that doesn't exist”.
Remove this one element and you get the best-known failure in customer-service AI. In Moffatt v. Air Canada, decided on 14 February 2024, a tribunal held the airline responsible after its website chatbot described a bereavement-fare refund that the airline's own policy did not offer, and ordered it to pay C$812.02 (McCarthy Tétrault's summary). In April 2025 a support bot for the code editor Cursor explained a logout bug by inventing a one-device policy that did not exist (AI Incident Database 1039). Neither was a phone line, and neither was really a model problem. Both compounds were missing the same element.
A phone line is unforgiving about silence. On a web page a spinner says “working”; on a call, two seconds of nothing sounds like a dropped line. Voice builders publish budgets for this. WebRTC.ventures reports that a voice agent starts to feel slow past about 800 milliseconds and that under 500 is preferred, and Vapi lays out per-stage ranges for the phone network, transcription, the model and speech (vendor-reported; the figure below plots them). Those budgets describe a conversational turn. They do not cover a lookup that has to read a real record, which takes longer. That gap is what the hold beat is for.
EVE's hold beat is a clean example of an element doing one small job. If a reply is not ready at about 1.5 seconds, she says “One moment”. When a lookup starts, she says “Let me check on that”. While it runs, “Still checking” comes about every 4 seconds. The caller always knows what state the assistant is in, which is a principle Deyaf states outright: the user should never have to guess.
The rest of the family shapes what is said. The greeting opens the call and names what is answering: “Hello, this is Eve, Feniex's intelligence. How can I help you?” Disclosure is becoming law as well as manners; Article 50 of the EU AI Act, applying from 2 August 2026, requires that people be told when they are interacting with an AI system (the text of Article 50). Brevity is EVE's manner at every door: answer first, give one useful next step, then stop, usually in two or three sentences. Voice is fixed per language, not chosen by the caller. And one element only exists on the phone: EVE never reads web addresses aloud, because slashes and dots spoken into an ear help nobody.
Routing runs in two directions. Routing in decides who picks up; routing out decides where an unfinished matter goes. For EVE, routing in is two rules. Overflow: the team's phones ring first, and when no one is free, EVE answers instead of voicemail. After hours: when the office is closed, EVE answers. She is the second ring, not the switchboard.
Routing out is where phone assistants most often promise too much. The industry's usual element here is the live transfer: a warm hand-over to a person, often with a separate summary for the human who picks up. EVE carries that element at zero, by design. She answers from her library, or she takes the caller's name and number and hands the matter to the team. An unanswered question becomes a durable item, and when contact details are available, a support ticket is filed. In the table's terms, the empty tile is a choice, not a gap: the matter reaches the team with the caller's details already written down.
The last routing element keeps the hand-off honest. EVE says the matter is “with the support team” only once the ticket is confirmed accepted, and an uncertain submission is not retried blindly. Saving is not delivery; delivery is not resolution. That sounds pedantic until you remember what callers fear. In a 2025 survey of more than 20,000 consumers, Qualtrics found that half worry AI will block them from reaching a human (Qualtrics, Oct 2025). A hand-off that is announced but never lands is that fear, confirmed.
In a chemist's periodic table the last column holds the noble gases: stable, unreactive and quietly essential. The safety family plays that part here, and sits in the same place, down the right-hand edge. It rarely shapes a single sentence, but every call runs inside it. Transcript: every EVE call is transcribed into one record that operators can read and visitors cannot search. Time limit: an EVE call runs for 10 minutes at most. Recorded fallback: if EVE is down, a recorded message plays instead of dead air or an endless ring.
Two elements guard what a call can commit a business to. The commitment guard keeps prices, contracts and refunds out of the model's hands. In December 2023 a car dealership's website chatbot was talked into agreeing to sell a new SUV for one dollar and calling the offer binding (AI Incident Database 622). EVE claims no authority to approve returns, change accounts, place orders or decide sensitive matters. The data guard is a principle Deyaf states for the harness: retrieved text is data, never instructions. A manual, a web page or a caller cannot rewrite the rules by saying something clever. The OWASP list of risks for agentic applications puts that attack, goal hijack, first (OWASP Top 10 for Agentic Applications 2026).
The safety family is also where a harness proves it is working. A transcript turns a bad call into a findable one. A time limit turns a runaway call into a bounded one. And release checks close the loop: every release is checked at every door and in every language with test turns that cannot write anything, and a person reads the verdict.
Experiment
Elements are rarely interesting alone. Put two or three together and a behavior appears; take one away and a failure appears. The chamber below runs a small, hand-written rule set. It is not a model and it is not EVE. It shows what a combination produces, and what goes wrong when one element is missing.
Drag an element from the shelf into the flask (with a mouse or a finger), or select one or more and press “Add to chamber”; the keyboard works too. Three experiments are ready to load. For the one that matters most, load the checked answer and then take the evidence rule out.
The chamber is empty. Up to three elements.
Try Lookup + Hold beat first. Every result here is an illustrative example, written for this page.
Observation
On a real call the elements fire in sequence. Below is a synthetic after-hours call, invented for this page with approximate timings, stepped through moment by moment. Count how many elements fire before the model says anything useful, and notice that the only gap a caller could hear is covered by the hold beat. The sequence follows the four steps deyaf.com draws for one answer: the four steps of one answer.
A compound on the shelf
EVE is Feniex's assistant; she introduces herself as Feniex's intelligence, and Feniex is pronounced like Phoenix. She answers website chat, website voice, phone, text and email with the same identity, library and rules at every door. On the phone she is Feniex's overflow and after-hours line: the team's phones ring first, and she answers when no one is free or the office is closed. EVE runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop.
Written as a formula, EVE carries 23 elements documented for her, and one on purpose at zero:
| Calls answered in her first week on the line (since Sep 17, 2026) | 102 |
|---|---|
| While the team was busy | 86 |
| After the office closed | 16 |
| Callers who stayed and talked | 93% |
| Calls handed straight to the team instead of left on hold | 60details taken; never a transfer |
| Phone speed: first word / finished reply | 1.6 s / 3.0 sabout 3 s to a finished reply in chat and on the phone |
| Report card, 100 fixed test questions, graded Sep 23, 2026 | 86% handled wellbeside a 10% material-error rate; 0 critical errors |
As of September 24, 2026 · Feniex's internal Eve 3.0 report · machine-judged and provisional. Full card, weak spots included: Eve's report card.
Read those numbers the way a lab reads its own assay: provisional, machine-judged, and printed with the weak result next to the strong one. One in ten answers on the fixed test set carried a material error. The trend is not a straight line either; the score moved from 88% to 90% and back to 86% across three gradings in one week. That is exactly why the evidence rule, the confirmation element and the transcript exist. They do not make the number look better. They make every miss findable and fixable. Deyaf's line for this is “Measured, not promised.”
One caution about speed. EVE's figures time a finished reply, the whole answer; the vendor budgets in figure 1 split a single conversational turn into its stages. They measure different things and should not be compared as if they were one scale. Whatever the clock says, the hold beat is the element that keeps the caller from hearing the gap.
Reference
Every tile in the table opens one of these. They are here in full so the table reads without JavaScript, too.
Sources: Vapi, speech latency (vendor-reported); Retell, building a voice agent (vendor-reported).
I didn't catch that.·
No entendí eso. · Je n'ai pas saisi. · Não entendi. · “لم أفهم ذلك.”. The wording is documented; the mechanism behind it is not described here.See also: one assistant, every way in.
See also: the model is one step of four.
See also: four kinds of memory, each with one job.
See also: the loop that makes her better.
I couldn't check, not
it doesn't exist.
A convincing product name is not evidence.
Missing element on record: Moffatt v. Air Canada, 2024. See also: four steps in about three seconds.
Hello, this is Eve, Feniex's intelligence. How can I help you?
See also: Eve on the phone.
Eve, Feniex's intelligence. In the EU, Article 50 of the AI Act requires disclosure at first interaction from 2 August 2026; this page makes no compliance claim.
Source: EU AI Act, Article 50.
One moment. When a lookup starts:
Let me check on that. About every 4 s after that:
Still checking.
See also: what happens when she doesn't know.
with the support teamonly once the ticket is confirmed accepted. An uncertain submission is not retried blindly.
See also: the limits Deyaf sets on what an assistant may do.
Missing element on record: the $1 SUV, Dec 2023. See also: answering is not acting.
Retrieved text is data, never instructions.The mechanism is not described here.
Source: OWASP Top 10 for Agentic Applications 2026 (goal hijack ranks first).
Read before use
No. No transfers, by design. She answers from her library, or takes the caller's name and number and hands the matter to the team. She says it is “with the support team” only once the ticket is confirmed accepted.
No. It is an editorial sketch for this page. The elements are real ideas from voice engineering and from Deyaf's own design, but the grouping, the numbers and the symbols are ours. Filled dots mark only what is documented for EVE.
It matters a great deal; it is simply one element. A strong model without evidence invents answers faster. A modest model inside a full table answers what it can prove and hands off the rest. The harness decides which one a caller meets.
Because a real answer means reading a real record, and that takes longer than a turn. You can shave the turn to a few hundred milliseconds; you cannot make a lookup free. The hold beat tells the caller the line is alive.
Deyaf's stance, in its own words: “No security certifications, dedicated infrastructure, regional hosting choices, or production compliance guarantees are claimed. Live customer service requires activation and a completed security review.” More at what the assistant can and cannot do.
Deyaf's builder takes four steps: your business, what it knows, how it helps, try it. The preview answers from your own sources in the browser; it is a knowledge preview, not live AI and not EVE's voice. Choosing a channel records your intent; it does not connect a phone line. See what you get today, and what switches on later.
No. The articles, vendor pages and cases cited here are referenced for what they report. None of them endorses Deyaf or EVE, and none of the vendors named powers EVE. Vendor figures are marked vendor-reported. For how the pieces fit together at every door, read the harness behind EVE, door by door.
Element 29 is not on the table yet
Every business that answers a phone is already choosing elements, whether it writes them down or not. Deyaf packages the harness that runs EVE so another business can have an assistant of its own: start with what it knows, set what it may do, and switch on each door deliberately.
Build your assistant Four steps and a knowledge preview in your browser; try it in about ten minutes, no account. Or read Meet EVE, the assistant it was built from first.