Agentic Harness: Press 1 is over.

Lecture notes · 1

Before any machine could understand a sentence, a business that wanted to sort its calls had one reliable signal to work with: the buttons on the caller’s phone. Each key on a touch-tone keypad sends its own pair of tones down the line, and equipment at the other end can tell which key was pressed without understanding a single word. That is the whole trick behind the phone tree, known in the trade as interactive voice response, or IVR. A recorded voice reads a list of choices, the caller answers with a digit, and the system follows a branch.

The phone tree solved a real problem. A switchboard operator could route calls by asking where the caller wanted to go, but operators were expensive and could take only one call at a time. A menu could pick up every call at once, day or night, and point each one toward the right desk. For a business with a sales team, a support team and a returns desk, “press 1 for sales” was a genuine improvement over one ringing line and a harried receptionist.

The trouble is structural, and it is worth writing on the board. A tree only knows what it was built to ask. Its designer has to guess, in advance, every reason anybody will ever call, then fold those reasons into a handful of branches short enough for a caller to hold in their head while a voice recites them. Everything the caller knows about their own problem, which is usually a great deal, is thrown away except for a single digit. And every level of the tree costs the caller twice: once to listen to the options, and again to decide which of them their problem most nearly resembles.

Callers learned to fight back. They pressed zero, said “agent” over and over, or chose any branch at all in the hope of reaching a person sooner. Designers answered with longer menus, the familiar “please listen carefully, as our options have changed”, and more levels. The tree kept growing because the business kept discovering new reasons people call, and a tree can only learn by growing. Hold on to that sentence; Lecture 3 is its answer.

Lecture notes · 2

The menu is only half of routing. The other half happens after the last press, in a piece of equipment the industry calls an automatic call distributor. Its job is to hold calls in a queue and hand each one to the next suitable person as they come free. Once contact centres had a queue, they could measure it: how many callers were waiting, for how long, and how many gave up before anyone answered.

Skills-based routing refined the queue. Instead of treating every agent as interchangeable, the distributor tags each person with skills: speaks Spanish, knows the warranty policy, handles large orders. A call that arrives carrying the right tags goes to someone with the matching skills, even if someone else has been free for longer. It is a sensible idea, and variations of it remain common in large contact centres.

Notice where the tags come from, though. The distributor cannot hear the caller. It knows the caller wants Spanish because they pressed 2, and it knows the matter is a warranty question because they pressed 4 and then 1. Skills-based routing is only as good as the guesses the caller made in the tree. If the person with a broken unit chose “technical support” because the unit is technical, they now sit in the wrong queue, and the fix is a transfer: whoever answers passes the call along, and the caller waits again.

In industry practice, transfers come in two kinds. A cold, or blind, transfer simply moves the call, and the next person starts from nothing. A warm transfer has the first person introduce the caller and summarise the problem before stepping away. Warm is kinder, but both share one weakness: the matter lives in the conversation, not on paper. If the line drops, if the next person is busy, or if the summary was rushed, the caller starts over. Repeating yourself is a familiar complaint in service: Zendesk’s CX Trends 2026 report found that 74% of consumers are frustrated when they have to repeat information (vendor-reported).

And then there is the overflow rule, the quiet last line of every routing plan: when no one is free, or the office is closed, where does the call go? For most businesses the honest answer has been voicemail, a recording that promises a call back and leaves the caller wondering whether anyone will ever listen to it.

Lecture notes · 3

Speech recognition gave the tree a new input. Instead of “press 2”, menus began to say “say ‘billing’ or ‘orders’”, which is still a menu, only spoken aloud. The bigger change came when systems could take an open question, “how can I help?”, and work out what the caller wanted from whatever they said. The industry calls this intent-based routing, or a conversational IVR. The caller speaks naturally, software sorts the request into an intent, and the call goes where the intent points.

That is a real improvement, and it is also where many projects stop. Classifying intent only replaces the keypad. The call still lands in the same queue, with the same wait and the same transfer at the end. Worse, a system that listens well but has nothing reliable behind it can sound confident while knowing nothing. The best-known chatbot failures of recent years were not routing failures; they were systems that answered when they should have looked something up, or handed over. In Moffatt v. Air Canada, decided in February 2024, a tribunal held the airline responsible for what its website chatbot told a grieving customer about bereavement fares, advice that contradicted the airline’s own policy page.

Here is what changes when the listener is part of an Agentic Harness rather than a classifier bolted onto a menu. Anthropic’s engineering guide to building agents names routing as one of the basic patterns: classify the input, then send it to a specialised next step. A harness does that too, but its next step is not only a queue. It can be the answer itself, drawn from the business’s own records, or a written hand-off to the people who can act. The listening and the answering belong to one system, working under one set of rules.

On the board, the difference fits in four verbs: listen, look it up, check the rules, reply. Deyaf puts it as the model is one step of four. Listening tells the harness what the caller means and which language they speak. The lookup finds the record that actually settles the question. The rules decide whether the assistant may answer, and what it must hand to a person instead. Only then is anything said. Routing stops being a question of which desk to ring, and becomes a question of where the answer lives.

Lecture notes · 4

Now the worked example. Eve is Feniex’s assistant, and at Feniex she answers the phone. Look carefully at where she sits in the routing plan, because it is the opposite of the usual pitch for AI on the phone. The team’s phones ring first. When a person is free, a person answers, and Eve never hears the call. Only when no one is free, or when the office is closed, does the call route to her, and she answers instead of voicemail.

In routing terms, Eve is the overflow rule. The slot that used to belong to a recording saying “please leave a message” now belongs to an assistant that can actually help. She greets every caller the same way:

“Hello, this is Eve, Feniex’s intelligence. How can I help you?”

That sentence does two jobs. It tells the caller they have reached an AI before anything else happens; disclosure at first contact is where regulation is heading, and since 2 August 2026 the EU AI Act’s Article 50 has required it of AI systems that interact with people in the EU. And it asks an open question, so there is no menu to learn. After the greeting, routing happens by listening. Callers who speak Spanish, French or Portuguese are answered in that language, in that language’s own voice; with English and Arabic, Eve works in five languages. Whatever the language, she searches one English library, so a caller in Portuguese gets the same facts as a caller in English. One assistant, every way in.

Her manner is part of the routing, too. She answers first, gives one useful next step, and stops. When she needs a detail, she asks for one at a time. Before any claim about a product, its compatibility, a warranty, a price or software, she checks the exact record. She never reads a web address aloud on a call. And she is never silent while she works. Voice engineers budget conversation in milliseconds; WebRTC.ventures’ latency guide notes that a voice agent starts to feel slow past about 800 ms, with under 500 ms preferred. A lookup can take longer than that, so the line fills the gap honestly: about a second and a half in, a short “One moment”; “Let me check on that” when a lookup starts; “Still checking” about every four seconds if it runs long.

A few rails come with the line. Every call is transcribed. A call lasts at most ten minutes. If Eve herself is down, a recorded fallback plays, so the line fails to a message rather than to silence. None of this is a separate phone product bolted onto a chatbot. Eve runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop. The phone is one more door into the same system. More on her: Eve on the phone.

Lecture notes · 6

Put the pieces together and you have the routing rules of a harness phone line. They fit on one board, and they are far shorter than a phone tree, because most of the tree’s branches have become lookups.

By hours

The first rule is about people, not the assistant. During open hours the team’s phones ring first. The assistant takes the overflow, the calls nobody is free for, and after the office closes it takes the calls that would otherwise have gone to voicemail. A business designing its own line should write down its hours and its ring order before it writes a word about AI; the assistant’s job is defined by the gaps in the team’s day, not the other way round.

By language

The old tree routed language with a key press and a queue of bilingual staff; if your language was not on the menu, you pressed 1 and struggled. A harness routes language by listening. Eve answers Spanish, French and Portuguese callers in that language’s own voice, searches one English library, and replies in the caller’s language: English inside, the customer’s language outside. Languages still deserve testing one at a time. The BabelArena benchmark, published this month, found that agents working in languages other than English used up to twice the tokens, made more tool errors, and slipped into English in structured output.

By topic

This is where the tree’s branches used to be. A harness does not route topics to desks first; it routes them to evidence. If the answer lives in the library, in a manual, a fitment record or a written policy, the assistant answers from it. If the question needs authority it does not have, or the record does not settle it, it takes the details and hands off. A failed lookup means “I couldn’t check”, not “it doesn’t exist”; unknown stays unknown. And every hand-off is a lesson waiting to happen: the answers the team gives can be reviewed and taught back, so the next caller with the same question gets it answered. That is the loop that makes her better, and it is supervised: only approved lessons are published.

By what is remembered

One ticket per matter depends on memory that is kept on purpose. The conversation record is one archive across web, email, text and phone, readable by the team and not searchable by visitors. Customer notes are short and historical, and never stand in for current facts. That separation is why a hand-off can carry the matter without carrying guesses: four kinds of memory, each with one job.