Custom Agent – Xedant

An AI employee designed and trained specifically for your business tasks, under your full control

Custom Agent is not just “your own ChatGPT”. It is a full agency-grade AI system installed on your server and deeply trained to solve your exact tasks. It doesn’t work “out of the box” — it learns under your control, and the further you go, the better it works and the less attention it needs.

Under the hood there is a powerful harness that turns a regular AI model into a reliable “employee”: with a set of skills, automatic checks, backups, and instant error fixing.

You can see how this looks live in the WordPress Agent service — there is a demo on a real site, so you can judge how fast and how precisely the system handles even the hardest tasks, and how you can retrain it if it ever falls short.

Business tasks we can solve

In theory, everything is possible. No jokes, no exaggeration: the models are already smart enough that, with proper training, they can solve any task. The real questions are access to the data they need, the time spent on training, and having the tools to execute the tasks.

Here I would split the tasks into two categories: the ones I have already solved successfully with agents. And the ones I am solving right now or planning to solve soon in my own business. If your task looks like any of these — write to me and get a fast result. If it doesn’t — write and ask: maybe it just hasn’t occurred to me yet, though it wouldn’t be that hard to solve.

Successfully solved tasks

Software product development

Managing websites, filling them with content

Building a management system for internal infrastructure

Auditing infrastructure for protection against attacks and errors

Visual analytics systems

Collecting and analyzing competitors and market trends

Collecting and analyzing potential advertising channels

Business analyst and advisor

Tasks in progress and planned

Managing a Telegram channel

Managing a YouTube channel

Managing business accounts on social networks

Finding and planning advertising in paid channels

Direct search for potential customers

Deep, automated market research

Tracking news and the market in real time

Managing tax reporting in a small business

Automatic improvement of existing agents

Monitoring and immediate reaction to incidents

Discuss your task

What is this, in plain words?

It is a professional service that develops a complete agent from scratch for the specific business tasks of the client. From idea to a result measured in solved problems or measurable profit.

This is NOT vibe-coding. NOT OpenClaw. NOT training and NOT coaching. It is professional development by a specialist with 25 years of building marketing and financial analytics systems — where data comes first: its quality, availability, protection, and safety.

The heart of the agent is the data. Most of our time will go into exactly that: collecting it, extracting it from existing systems, validating, cleaning, and organizing it in a form the agent works with well

You don’t do this by hand — but for this task we will need to train the agent: to understand what you have and how it can be used, how to handle the data properly, and what can be built on top of it

Once the data is in place, it becomes easy and pretty to layer interactive things on top: bots, reports, actions in the real world

Second most important in business tasks — replacing (where possible) the model’s hallucinations with code (scripts). Instead of counting “in its head”, the model learns to write, debug, save, and reuse small focused scripts for specific tasks

Scripts combined with quality data are exactly what guarantees the absence of model “hallucinations” and bad decisions — especially when it performs real actions

Even if you don’t ask the model to write scripts, it writes them anyway — one-off and hidden from you. That is exactly what causes errors in answers and heavy token spending in systems that have no scripts of their own

When the data and scripts are ready, and the system has learned to use them — and to extend and fix them automatically — you can create freely, for the good and the growing profit of your business

Anyone who has already tried vibe-coding on real business data will understand exactly what I mean here 🙂

At first the AI is spot-on, then the fantasies begin — and good luck noticing them. It’s like your trusted employee went mad and started raving. And your arms drop, because you’d have to teach it everything all over again — a recent client, about their past experience with ChatGPT on the same tasks

What’s great in my case is that the product data lives in a database. It updates. Yes, I noticed the AI counting it wrong. It laid out for me right away where each number came from. I pointed at the error. It fixed the script and recalculated the whole database. That is damn hard to get anywhere else. Especially with a hundred-plus products. You feel something is off in the conclusions. And the AI just says: “too many products, I decided to simply analyze the first 10” — a recent client, about their current experience with their personal agent

How this differs from ChatGPT and the like

You have probably tried ChatGPT, Claude, DeepSeek, or something similar. The result was likely this: great at generating text, but weak at solving specific tasks, ignorant of your data, and impossible to retrain.

Your agent is fundamentally different:

Under the hood is a full agency system, not just a chat. It executes complex tasks while continuously gathering deep knowledge about your business — knowledge no public chat or service has

When ChatGPT “acts dumb” — all you can do is swear quietly. In your own agent you can fix the behavior immediately with one short message about what’s wrong and what you actually want

The system is installed on your server — your data never leaves it, full isolation

The skill set — text instructions the system receives on every request. Not “the model’s built-in memory”, but your controlled and documented knowledge base

Automatic checks after every action — the system finds and fixes its own mistakes

Backups, versioning, and recovery — the skills and data are protected

Works with any AI models — Claude, OpenAI, DeepSeek, Qwen, GLM, Kimi and more, local models, OpenRouter

Access through a web chat or Telegram — from wherever is convenient

Analytics — statistics for every action: tokens, cost, speed

Proven on real tasks — from managing websites to marketing research, managing ad campaigns, product development, deep analytics, and machine learning

Even as models inevitably change over time, your personal assistant only gets smarter thanks to the accumulated skills

How this differs from OpenClaw, Hermes, Claude Code, OpenCode, Codex, and the like

Vibe-coding tools: Claude Code, OpenCode, Codex

Claude Code, OpenCode, Codex, and others are vibe-coding systems built specifically for application development. They work well for developers who understand exactly what they want to get and can supervise the execution (no complaints here — I’m one of them). Or for the vibe-coders who build “something” at their own risk, without understanding the real internals of the interfaces they see.

That second group is a lottery: it may turn out pretty, but with hidden and serious risks of data loss or leaks — which can be costly and painful for any business. And the development process itself isn’t always pleasant: when you don’t understand what’s happening, you can spend hours typing “the same thing again”, and the model will dutifully fix one bug while adding two more and breaking what worked perfectly before.

“Virtual assistants”: OpenClaw, Hermes

OpenClaw, Hermes, and others are “virtual assistants” — a universal wrapper around language models with memory. In theory, they should adapt to the user and learn on the fly. In practice, getting stable work out of them is even harder than out of vibe-coding tools. Why? Because in a vibe-coding system you at least see what’s happening “under the hood” and can edit the instruction and skill files. Here you face a black box: yesterday it worked well, but today it decided to take your recent wishes and chats into account, changed its internal rules, and suddenly got dumber.

Asking it “please note that this job has to be done like so…” is like handing a grenade to a monkey: pure luck. It may take it into account and do it right, it may ignore you, it may even break what used to work. And one more important nuance of such “assistants”: they consume tokens constantly and very wastefully. Where a vibe-coding tool’s limit would last a week, OpenClaw can burn through it in a couple of hours. Personal experience, for what it’s worth.

“Homemade” harnesses: Pi, Aider, and others

“Homemade” harnesses like Pi, Aider, or any of the countless lone-inventor creations — they’re pleasant to write for yourself, but next to impossible to use for serious tasks. Instead of solving real problems, you endlessly solve technical ones, heroically overcoming the limitations of your own “homemade bicycle”. It’s especially hard when the bicycle is someone else’s: not only is it full of bugs, but the developer answers, at best, once a week or even once a month. Given the flood of vibe-coded pull requests in open source these days, many authors have started simply deleting both support requests and proposed bug fixes — demanding proof that the request and the proposed code were written, or at least verified, by humans.

My approach

I was solving concrete online-business tasks from the very beginning, so I tried to take the best of both worlds: instead of writing my own harness from scratch, I simply “wrapped” Claude Code. Today it is effectively the strongest harness in existence, and it adapts easily to any models from any provider — not necessarily Anthropic’s. But I went much further than a web interface on top of it: iteratively, over a year, I built a whole infrastructure above it that makes it possible to create truly autonomous and easily trainable agents. I turned a developer’s command-line tool into an infrastructure for an agency system. I can even swap Claude Code for any other harness with any other models if I want — right now there’s simply no need: my system runs stably and needs no such experiments.

This is NOT a vibe-coding tool. If vibe-coding is what you want, it will honestly be easier for you to take any product from the list above.

This is NOT a “virtual assistant” like OpenClaw. If a “helper” is what you want, it will again be easier to take and configure any existing off-the-shelf solution.

This is NOT a “homemade harness” like Pi. It is a full agency system built for business, where Claude Code (currently the most “grown-up” harness) works as just one of the building blocks — and an easily replaceable one if needed. I have no desire to write my own bicycles.

With me you get the best of both worlds: you can talk to the agent as if it were a living person, while “under the hood” the agent writes large volumes of code and works with real data. That’s why its answers and actions are always fast, precise, and cheap. On one side, it’s a software product and a system for developing a software product. But that complexity is hidden from you behind the “living” conversation and the ability to retrain on the fly. When you ask it to take something into account for the future — the system doesn’t just add a rule, it can update dozens of scripts along the way, with fully automatic checks and error fixing.

Because the system turns any task into internal scripts, it is genuinely faster and genuinely cheaper — both to build and to run. When you ask for “a sales report for the last quarter” somewhere in OpenClaw, it will go and write the very same scripts anyway — just on the fly. It will make mistakes and fix them on the fly at the cost of your time and token limits. The next chat starts everything over. In my system, the script is written once, at your first request; after that the system knows when and how to call it, and can fix and extend it on the fly, invisibly to you.

The right way to automate your tasks

The main problem is not generation. Generation is cheap and practically unlimited now. The main problem is control. Without control and retraining, the system works poorly and unstably — it generates a lot, but of low quality. This is exactly where the disappointment with AI comes from: it looks like what you asked for, but it’s still “not quite it”.

The best results come not to those who are experts in AI or development, but to those who are experts in the specific problem being solved — they are the ones who can adequately control and retrain the system

Being a technical specialist can even be harmful — they are used to saying “what” to do and “how” to do it, robbing the system of the chance to show itself in full

It works better to ask for results, for goals — not “what to do” but “what I want to get in the end”

Never “forgive” the system even small mistakes — better to ask once for it to fix them and learn. It costs nothing more than a simple request in a free-form chat

The main work is not installation and setup, but feeding the system fresh data, supervising and retraining it, keeping it stable and safe

You need to unlearn the habit of doing everything by hand, even if it’s “just one line to fix”. In practice it’s simpler and faster to ask the system to fix something, quickly retraining it if it errs. You’ll save time and energy in the end

Appetite comes with eating. Experiment more — especially with the tasks that now seem too hard or downright impossible for AI. That’s exactly where you’ll get the maximum benefit and the competitive edge

Skills — the main asset

The most valuable part of the whole system is the skills. They are a set of text instructions the system receives on every request. The skills are what make the system smart — and they are what takes longest to restore if lost.

Skills are just small text files — easy to copy and store

Skills created for one task often fit any similar task — with cosmetic edits

Your personal skill of growing the system’s skills will be priceless — that is exactly what defines the competitive edge of the future

Full change history of the skills is stored in git — you can always roll back to any past version. But it’s still better to keep many copies in different places — I’ll show you how, automatically and without effort

If you don’t know what git is — it’s the world standard for keeping the full change history of text files, with fully automatic synchronization between computers and practically any electronic devices

Scaling

Once you’ve figured out how the system works, you’ll realize it can do more than solve the current task — it can set up its own copies for any other tasks, under your control.

This is actually not hard. But it helps to have a technical specialist at hand — to consult the system itself on the available scaling mechanisms and to bound its flight of fancy in the process. The technical specialist is not there to do everything for you — but to back you up and show the boundaries.

Skills from one task carry over to similar tasks with minimal edits

One system — any number of tasks and projects

Skills are a permanent asset: copy them, scale them, pass them on

Autonomous licensing — if the author disappears, nothing happens to your projects

Installation, setup, and training

The process has three stages: buying the necessary accounts, installation and setup, training.

Our work starts with you buying a server and access to language models. You buy the VPS server and the model subscription (OpenCode Go or Command Code GOAT — $9/month) under your own data — that’s safer and simpler in the long run. I’ll help you pick suitable options and recommend proven providers and payment methods. Details on the necessary expenses are in the Expenses section at the end of the page.

Installation and setup are done by me personally after you grant access to the VPS you bought — neither you nor your team needs to take part at this stage.

Training is joint: you know your task better than I do, and I know my system and its quirks better than you do. The further we go, the less work for both of us: the system gets more stable and smarter with every request — not just every day. Within 1–2 weeks I won’t be needed at all, except occasionally, for a couple of minutes.

Buying the necessary accounts — 1 hour. You buy the accounts yourself, under your own data — that’s safer and simpler in the long run. I’ll help with picking suitable options and finding payment methods. Details in the Expenses section

Installation and setup — 1 day. You grant access to the purchased VPS, I install and configure the system, connect the AI model, and create the base skills: initial documentation, rules of work, data protection

Training the base skills — 1 week. You set the tasks, the system executes, we check the results together and retrain. If the system errs, you write what’s wrong, I make sure it turns into a corrected skill and give practical advice along the way. Within days the system learns enough to work confidently and stably — with occasional slips, but the typical tasks it performs practically error-free

The further we go, the less my help is needed. The system copes better and better, and you can already retrain it on your own, without me. My role at this stage is technical backup and help with expansion.

Can you do this with your own hands?

Theoretically, yes, it’s possible. Register and buy the model subscription yourself (OpenCode Go or Command Code GOAT), install and configure the harness, write and debug the first skills and validation scripts, arrange safe access to data, preservation and accumulation of the models’ experience. Plus hundreds of moments I could talk about for hours based on my own experience. It is, of course, an interesting experience. But the main question is — do you want to grow your business or your technical skills? Will it be a toy or a tool?

I started with applied AI about 3 years ago, and before that I had been programming for more than 25 years, 8–12 hours a day, no days off. And my harness has been written, used, and debugged for over a year across almost 50 projects. For your business-automation result to be at least “no worse”, you’d want experience like mine. Otherwise you’ll drown in little things that have little to do with AI — but without them everything else breaks. If not today, then tomorrow — inevitably.

This also answers the question of why I sometimes insist on this or that approach or tool. It’s usually the result of my own painful experience. The price of a mistake is days and weeks of lost time — and if you go for the popular quick-and-dirty vibe-coding, add the risks of data loss, leaks of customer information, and legal trouble from small technical errors along the way.

So yes: you can play developer. Just don’t treat it as a way to solve real business tasks. Prototyping ideas is fine — but the implementation is better handed to people who can spot and fix the problems that aren’t obvious to you. To repeat: the main problem is not generation, it’s control. You can supervise the business logic, but you’re unlikely to notice the important technical nuances.

And if you like the idea of doing it yourself but don’t want to figure it out alone — there is a separate service for that, DIY Agent: I install the system on your server, set it up, and walk you through it for two weeks — until you solve complex tasks fully on your own. It’s cheaper than a custom build, because you do the main work.

What you are really paying for

When you order this service, you’re not really paying for the harness, or even for installing and configuring it. You’re paying for my experience of working with business data the right way — 25+ years of building business products, with thousands of clients and very large volumes of data.

It’s not enough to tell the model “here’s my database, here are my requirements, go do it”: in practice there are always many nuances of data formats, access, security, backups, and the right way to organize the workflow around them

I can spot in time where the model makes mistakes critical for working with data, correct it, and retrain it on the fly so you never hit that again

I can help arrange access to data where it’s technically hard — where there is no public documentation or even open access yet

I can structure data correctly for different tasks — from analytics to real-time decision systems

My services are especially valuable in tasks involving marketing analytics and financial analytics — those are the areas I always prioritize

Security will be top-notch: I always control model access to data strictly, and propose and implement solutions for protecting and anonymizing personal data, so you never run into legal trouble

I teach you how to work with models the right way — how to control and retrain them: gathering the same experience yourself would take months of trial and error

This doesn’t mean that I — or even my products — become irreplaceable. Once you understand, with my help, how the system works and how to work with it properly, you can migrate to other solutions at any moment, or even write your own, if that’s your thing.

My goal here is not to tie you to me — quite the opposite: to make you independent and self-sufficient as fast as possible.

Expenses

To get started there are a few unavoidable expenses — without them the work would be impossible or much harder and slower. None of these expenses are payment to me for the service: you buy them yourself, under your own accounts. That’s safer and simpler in the long run.

VPS server — from $6/month. Minimum 2 cores and 4 GB of memory for Docker, the harness, and the auxiliary data tools. It can go on your existing server, but then the risks appear right away: “what if I break something” or “what if its configuration isn’t enough”. Older servers, for example, may not support AVX2 — in that case only older harness versions are available, slower and worse at executing tasks. Once the system is fully set up and doing your tasks — I can move it to your infrastructure, if you like. But for the first month it’s better on a fresh dedicated VPS from a provider you trust

Model subscription — $9/month: OpenCode Go (about 2 billion tokens) or Command Code GOAT (about 4 billion tokens). More than enough for the first month; access to all the models is included in the subscription. I’ll help with registration and picking payment methods

You do NOT need to pay for Claude Code or Anthropic models. The model subscription — OpenCode Go or Command Code GOAT — already includes access to all the models: no “extra” payments, no registrations anywhere else

~3–4 hours of your time daily in the first few days of work — not constantly, but in bursts. Together we’ll pick and implement the automation for your business tasks, and I’ll teach you how to use and retrain the system properly. The further we go, the less of your time and mine it takes. The system will do more and more of the work, without you or me

Let’s discuss your ideas

Just write to me and tell me which task you want to solve. No preliminary contracts or prepayments for consulting.

Development costs $597 for two weeks. That time is usually enough to bring the agent to the level where you can keep improving it yourself, in chat mode. If my help is needed beyond that — $190 for each following week.

To estimate the timeline, don’t scare yourself with past custom-development projects. There is a “boxed” part here already: installing and configuring it takes one day at most — your direct help isn’t needed at that stage, I do it all after you grant access to the purchased VPS. After that it turns into a process not much harder than supervising and training a junior remote employee, in chat mode.