Extra: The Harness Era

Start with what a model is by itself: a fluent writer that has read a great deal and can produce plausible sentences about anything, which is precisely the danger. Ask it about your warranty terms and it will write something that sounds like warranty terms.

A harness turns that writer into a reliable employee, and the easiest way to picture one is the newsroom that printed this page. A paper does not run whatever its best writer types. The writer gets the file from the library; the copy desk checks every name and number against the record; the standards desk decides what the paper may say; circulation decides where the edition goes; and the corrections column makes sure a mistake is fixed once, in public, and not repeated.

An Agentic Harness has the same departments. Knowledge is the library: the exact records and reviewed lessons the assistant may use. Memory is the morgue, the newspaper’s archive, with the same rule that an old clipping is history, not today’s fact. The doors are circulation. Permissions are the standards desk. The learning loop is the corrections desk. Deyaf compresses one answer into four steps: listen, look it up, check the rules, reply. The model is one step of four.

Practitioners have their own vocabularies. Böckeler sorts a harness into guides, which steer an agent before it acts, and sensors, which catch problems after; some are plain code, some are another model’s judgment. Anthropic’s engineers note that every harness component encodes an assumption about what the model cannot yet do, and that those assumptions go stale as models improve, so a good harness is pruned as well as built. Philipp Schmid calls the model the CPU and the harness its operating system.

The vocabularies differ; the lesson does not. Reliability is not a property of the model. It is designed around it.

For a while the fear was being locked into a model. By September the argument had flipped. On September 1, Kai Waehner wrote that “the model is a config setting and the harness is your architecture.” He walked through the layers around the model, from tool connectivity to the runtime, named context and memory the stickiest by far, and proposed exit questions for every buyer: can you export your definitions, where does your context live, and who owns the audit trail?

Databricks made the same point from the builder’s side. “The core agent loop is just 1% of the work,” it said at its June summit; the rest is capacity, deployment, security, evaluation, monitoring, context and sharing. Days earlier it had introduced Omnigent, an open “meta-harness” that sits above other harnesses to compose and govern them.

The model is a config setting and the harness is your architecture.Kai Waehner, September 1, 2026

For a small or mid-sized business, the reading is simpler. The assets that matter are the ones you would rebuild if you switched: approved knowledge, rules, the record of conversations, taught lessons, and the decisions about which door does what. Those should be yours and legible, whatever model sits in the middle.

That is why Deyaf treats a customer-owned data boundary as a principle, and why it suggests you start with one good job. A harness that begins narrow is easier to understand and easier to measure.

The summer’s big technical story was about plumbing. On July 28 the Model Context Protocol, the open standard many agents use to reach tools and data, published a major revision. The core became stateless, confirmation can happen mid-request, and long-running tasks became an extension. Its TypeScript and Python SDKs had each passed a billion downloads.

The governance changed hands earlier. In December 2025 the Linux Foundation formed the Agentic AI Foundation to hold MCP, the goose framework and the AGENTS.md convention under neutral governance, counting more than 10,000 published MCP servers at launch.

Why should a harness reader care? A shared protocol makes tools portable between harnesses, the least sticky layer on the Business page. The decisions about what an assistant may touch, and on whose authority, stay in the harness. NIST’s February initiative makes the same point from the security side: agents need their own scoped identities, not borrowed service accounts.

One of the year’s harness stories is not in a lab. It is on a phone line. At Feniex (say it like “Phoenix”), the team’s phones ring first. When everyone is on another call, or the office has closed, the call goes to Eve, who opens the same way every time: “Hello, this is Eve, Feniex’s intelligence. How can I help you?”

Eve is Feniex’s warm, composed public face: a knowledgeable customer desk that helps people choose suitable equipment, make a buying decision, or get support for equipment they own. She is software with a name, not a person, with no age and no pretense otherwise. She answers first, gives one useful next step, then stops. She asks for one missing detail at a time and never tacks a sales pitch onto a support answer.

On the phone, the harness does work the caller never sees. Dead air sounds like a dropped call, so she is designed never to go silent: “One moment” if the line goes quiet for about a second and a half, “Let me check on that” as a lookup begins, “Still checking” every few seconds after. Every call is transcribed; a recorded fallback plays if she is ever down; calls run ten minutes at most. And there are no transfers, by design. She answers from her library, or takes the caller’s name and number and hands the matter to the team. For an emergency, she tells people to call 911.

The same assistant speaks English, Spanish, French, Portuguese and Arabic, from one library, not five. A question in Spanish is rendered into English to search the single library, and the answer returns in Spanish, in that language’s own fixed voice. Detection and translation run separately from the main answering model.

What she may do is deliberately small: look something up, search the taught library, take a message, record an unanswered question. The tool list and server-side queries enforce that scope. Before any exact claim about a product, compatibility, warranty, price or software, she checks the record, because a convincing product name is not evidence.

When she cannot answer, the question does not vanish. It becomes a durable unanswered item and, with contact details, a ticket. A person answers it once; it is checked, then taught. She is measured on a frozen test set before anything is taught, and again after: a supervised loop, not fine-tuning. More about her, from whether she is a person to what you can build in about ten minutes, is in the FAQ at the bottom of Meet Eve.

Most people meet Eve in one of two places. The first is the chat box in the corner of a Feniex web page. Her reply streams in as it is written; the page and the product the visitor is looking at ride along with the question; a product or resource tile can sit beside the answer; and she keeps a bounded history of this conversation only. Website voice works the same way, out loud.

The manners do not change with the door. She answers first, offers one useful next step, then stops, usually in two or three sentences, and asks for one missing detail at a time. Before any product, compatibility, warranty, price or software claim she checks the exact record, because a convincing product name is not evidence. She never pretends to be a person.

The second place is the phone. The team picks up first; when no one is free, or the office has closed, Eve answers instead of voicemail. Routing in is overflow: people first, then Eve. Routing out is a hand-off of details, never a transfer: she answers from her library, or takes a name and number for the team, and she never reads web addresses aloud. Callers who speak Spanish, French or Portuguese are answered in that language’s own voice.

At either door, a question she cannot answer becomes a durable unanswered item and, with contact details, a ticket. She says “with the support team” only once that ticket is confirmed accepted. The phone side is told at length in Eve on the phone.