Systems essay 04 10 min read · Updated Sep 2026

The work lives around the model.

That's the Agentic Harness.

A model can reason. A production system must also remember, choose tools, recover from failure, show its work, and know when to stop. The Agentic Harness is the engineered runtime that makes those behaviors repeatable.

See a harness at work: meet Eve Read the field guide

Production runtime / v1governed

Business task arrivesinput 001

intent router

01Identity + instructionsoperating posture

02Context + memoryrelevant, durable

03Reasoning modeldecides next step

04Scoped tools + policiesprivileged interfaces

05State + verificationcheckpoint, test

06Traces + evaluationobserve, improve

Consequential action → explicit approval boundary

Editorial starting point. This original field guide builds on the production architecture described by Mrunmayee Rane in “The Agentic Harness: How to Build AI Agents in Production”. It extends that technical frame into a business-readable system map.

01The structure

What the harness surrounds

The model is one layer. The harness is seven more.

The source article makes a useful architectural move: responsibilities that are often crammed into a prompt are externalized into owned layers. That makes the intelligence easier to control, inspect, replace, and improve. Deyaf’s definition says it in one sentence: “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.” How an Agentic Harness works

Identity

Defines the agent’s role, priorities, communication style, boundaries, and stop conditions.

Task contract

States how work begins, what “done” means, how progress is checked, and how risk is reported.

Context

Routes fresh, relevant working information to the model without treating the context window as storage.

The model is one step of four

Memory

Preserves durable facts and decisions outside the conversation so long-running work can continue.

Four kinds of memory, one job each

Reasoning

The model interprets the task, chooses a next step, and adapts when a rigid path cannot.

Tools + skills

Tools expose controlled actions. Skills encode reusable procedures. Policies determine permitted use.

Answering is not acting

State + recovery

Checkpoints, retry rules, progress artifacts, and escalation paths prevent silent failure and lost work.

Observability, evaluation + governance

Traces reveal the trajectory. Evals grade both outcome and process. Governance gives permissions, changes, and incidents a clear owner.

A dated report card, weak spots included Deyaf’s trust and control principles

02How it runs

The production loop

A useful answer is not enough. The path must hold up.

Production quality comes from the execution loop living outside the model. The harness supplies continuity, policies, and evidence at each transition.

01 / receive

Frame the task

Capture the request, outcome, user identity, risk, and scope.

02 / load

Assemble context

Retrieve the right instructions, memory, sources, and procedures.

03 / decide

Choose the path

Route simple work to code; flexible work to an agent loop.

04 / act

Use a tool

Execute within permissions, budgets, retry limits, and approval rules.

05 / prove

Verify progress

Inspect results, test the outcome, and recover or stop when needed.

06 / learn

Persist + trace

Write durable state, record the trajectory, and feed evaluation.

The model reasons · the harness routes, constrains, remembers, verifies, and records

03System types

Choose the simplest shape

Not every AI feature should become an agent.

A healthy architecture matches agency to uncertainty. The more freedom a system has to choose its path, the more policy, observation, evaluation, and human control it needs around that freedom.

User requesttask + identity + risk

→

Known pathdeterministic workflow

Uncertain pathagent loop

High-risk pathhuman review gate

Type 01Agency / none

Single model call

One bounded input produces one bounded output. Fast, economical, and easy to test.

Best fitClassification, extraction, rewriting, structured summaries

Type 02Agency / low

Retrieval system

The system fetches relevant knowledge before the model answers, with sources available for inspection.

Best fitPolicy Q&A, knowledge search, grounded support

Type 03Agency / fixed

Deterministic workflow

Code owns the sequence. AI can help inside a step, but it does not invent the process.

Best fitRepeatable operations, compliance-heavy flows, known decisions

Type 04Agency / gated

Human-in-the-loop

The system researches or prepares; a named person reviews consequential action before execution.

Best fitCustomer messages, financial changes, hiring, publishing

Type 05Agency / adaptive

Single agent

The model chooses steps and tools across a bounded loop, adapting as evidence changes.

Best fitInvestigation, coding, research, exception handling

Type 06Agency / coordinated

Multi-agent system

Specialized agents work in parallel or review one another under explicit handoff contracts.

Best fitBroad research, independent verification, large system analysis

An industry map of system shapes, not a product ladder. Deyaf’s live product publishes three levels of autonomy instead: answer only, prepare for approval, and run an approved task. Each task and each channel moves up separately, and there is no single switch that lets an assistant do everything. See the three levels.

04Reliability

Production is a reliability problem

Measure the journey, not just the destination.

A polished answer can hide a bad trajectory: the wrong tool, stale context, needless retries, unsafe edits, or an unverified result. The harness makes that invisible work inspectable.

PassTask success and consistency

TraceTool and state trajectory

CostTokens, tools, time per success

RecoverFailure and escalation behavior

What to measure, defined, not scored. For one dated set of real numbers, see Eve at Feniex.

Observability

Capture context, tool calls, outputs, state transitions, retries, latency, cost, and feedback.

Trajectory evaluation

Grade how the work was done—not only whether the final response looks correct.

Failure policy

Define retry limits, safe alternatives, escalation points, and conditions that require a stop.

Governance

Give identities, permissions, memories, skills, releases, and incidents a clear owner.

Human control

Keep consequential writes visible, explainable, and gated by the right person.

05In production

The harness, already answering

The layers are not theory. They already answer.

Eve is Feniex’s assistant, and the clearest way to watch this guide’s map at work. Eve runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop. Meet Eve, Feniex’s assistant

Flat poster illustration of a small desk with a lime lamp at the center of an open, roofless room, with thin orange lines running from the desk to every doorway in its walls. — One desk behind every door. Illustration for mood only.

Eve at Feniex / one questionillustrative

“Does my warranty cover this?”any door

same library, every door

01Identity + ruleswho she is, what she may say

02Exact reference lookupthe record, checked first

03Reasoning modelanswers, then stops

04Scoped actiontake a message · record the question

05Hand-off to the teamunresolved → a person

06Conversation recordkept for the team

Consequential action → a person says yes

Illustrative example: not a live Eve transcript. One question walking down the layers of this guide’s map.

01 / Who

Who she is

Feniex’s warm, composed public face (Feniex is said like “Phoenix”). She helps people choose suitable equipment, make a buying decision, or get support for equipment they own. She is software with a named personality, not a person, and approved Feniex operators manage her rules and knowledge.

02 / Doors

Where customers reach her

One Eve, every door: website chat, website voice, phone, text and email, in English, Spanish, French, Portuguese and Arabic. On the phone she is the 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.

One assistant, every way in

03 / Knows

What she knows

One library behind every answer: versioned catalog, manual, software, fitment and policy records, plus lessons the team has reviewed. She checks the applicable record before any exact product, warranty, price or compatibility claim, because a convincing product name is not evidence.

Inside one answer, step by step

04 / May do

What she may do

Four public actions: look something up, search the taught library, take a message, and record an unanswered question. She describes only what the result proves (saving is not delivery, delivery is not resolution) and claims no authority to approve returns, change accounts or place orders.

What she may do, level by level

05 / Unknown

When she doesn’t know

Unknown stays unknown. The question becomes a durable unanswered item and, when contact details are available, a support ticket. She says it is with the support team only once the ticket is confirmed. No transfers, by design: on a call she takes the caller’s name and number and hands the matter to the team.

What happens when she doesn’t know

06 / Memory

What she remembers

Four kinds of memory, each with one job: this conversation, one conversation record the team can read, short customer notes, and her reviewed library. Orders, invoices and balances are always looked up fresh, never recalled from notes. Raw conversations never become public knowledge on their own; only lessons a person approves are taught.

How her memory is split The loop that makes her better

Report card86%handled well on 100 fixed test questions, graded Sep 23, 2026, with a 10% material-error rate and 0 critical errors.
Eve’s full report card

Phone102calls answered in her first week on the line: 86 while the team was busy, 16 after the office closed. Sep 17–24, 2026.
Eve on the phone

Hand-offs0 lostSep 16–23, 2026. Every one sent reached the team.

As of September 24, 2026 · Feniex’s internal Eve 3.0 report · machine-judged and provisional. These are Feniex’s numbers, not a forecast: a new assistant starts fresh with its own knowledge.

06At the door

Where customers meet the harness

Two doors customers see. One harness behind them.

Everything in this guide runs out of sight. Customers meet it in two places: a small chat box on a web page, and a phone line that rings when the team can’t pick up. The same identity, library and rules stand behind both; only the delivery changes.

~3 sTo a finished reply, in chat and on the phone

93%Of callers stayed and talked

60Calls handed straight to the team, details taken

10 minThe longest a call runs

Calls since Sep 17, 2026, her first week on the line. As of September 24, 2026 · Feniex’s internal Eve 3.0 report · machine-judged and provisional. Eve’s report card, speed included

The chat box

The smallest, most visible door. Replies stream in as they are written, the page and product in view ride along, an answer can carry a structured product or resource tile, and history stays inside this conversation.

The Box in the Corner: a chat box, as a comic From Script to Source: a history of chat

Chat manners

She answers first, gives one useful next step, then stops, and asks for one missing detail at a time. The exact record comes before any product, warranty, price or software claim. She never pretends to be a person.

Mind Your Manners: chat-box etiquette

The phone 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? Callers in Spanish, French or Portuguese hear that language’s own voice.

Nobody Has to Hold: automatic call taking

Never silent

“One moment” at about 1.5 s, “Let me check on that” as a lookup starts, “Still checking” about every 4 s. Every call is transcribed, a recorded fallback plays if she is down, and she never reads a web address aloud.

The Elements of a Phone Assistant, as a periodic table

Routing, both ways

No transfers, by design. In: overflow, the team first, then Eve. Out: she answers from her library, or takes the caller’s name and number and hands the matter to the team. In chat, she says “with the support team” only once the ticket is confirmed accepted.

Press 1 Is Over: phone trees, queues and intent routing Same rules at every door

DEYAF.

From a working assistant to yours

Built from Eve. Packaged for your business.

Deyaf is built from Eve: the harness that runs Eve at Feniex, packaged so another business can have an assistant of its own. 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 (browser-local, not live AI). Live doors switch on one at a time, each one deliberately.

See the harness behind Eve

Eve at FeniexAnswering Feniex’s customers in production today.
Watch Eve answer on five doors

Five doorsWebsite chat, voice, phone, text and email, with one library behind them.
Five doors, one library

Five languagesEnglish, Spanish, French, Portuguese and Arabic, answered from one English library.
How Eve handles five languages

Name one outcome

Choose a bounded, measurable business result with a responsible owner and an honest baseline.

Choose the system type

Use a model call, retrieval, or fixed workflow when it is enough. Add an agent only when adaptation pays.

Map context + authority

Decide what the system may know, where truth lives, who can access it, and what must never enter context.

Constrain tools

Give every tool an allowed scope, preconditions, failure behavior, retry limit, logging rule, and approval boundary.

Instrument the loop

Trace decisions and state changes. Evaluate repeat runs for success, consistency, cost, safety, and recovery. Measure on a fixed test set before anything new is taught. Measured before anything is taught

Widen on evidence

Promote capability only after the system repeatedly earns trust under realistic conditions. Switch on one door, one task at a time. Switch on each door deliberately

08Questions

Questions worth asking.

Is an Agentic Harness a new model?

No. The model supplies reasoning. The harness is the runtime around it: identity, context, memory, tools, state, policies, verification, traces, evaluation, and governance.

Does every AI system need an agent?

No. Many useful systems should remain a single model call, a retrieval flow, or a deterministic workflow. Agency is valuable when the path cannot be fully known in advance.

What is the difference between memory and a skill?

Memory stores durable facts—what the system knows. A skill stores a reusable procedure—how the system performs a kind of work.

Why is a human gate still important?

Reading and preparing create value with limited operational risk. Consequential writes can affect customers, money, people, or records. A named approval boundary keeps responsibility visible.

Where does Deyaf fit?

Deyaf packages the harness that already runs Eve at Feniex, so a business can set up an assistant of its own: its knowledge, its rules, and the doors it chooses to open, one at a time. The starting plan requires human review before external messages, record changes, or commitments, and every task begins at the first of three levels: answering is not acting.

Build the system around the intelligence.

Deyaf’s builder walks through four steps and ends in an exact-source knowledge preview (browser-local, not live AI). Try it in about ten minutes, no account needed.

Build your assistant Or start with one good job