An agent that works the way your company works.
A generic agent knows your tools. It has never been told how your company works, so nobody can hold it to your definition of right. We build one standing agent for one company, deploy it into that company’s own accounts, and keep it right after it is live.
- Spec and fixed price
- Within one working day
- Ready-made agent live
- In five working days
- Who owns it
- Your accounts, your keys, your code
- Channel
- #paymentswhere the finance team already works
- Authority
- Acts behind an approval gate
- Runs in
- Your own Cloudflare account
- Nadia
@atlas summarise last night’s failed payments and draft the follow-up to each customer
- @atlas
Nine payments failed after 21:00, four of them over €500. Nine drafts are ready in your outbox.
Sending reaches your customers, so it needs a name on it.
Companies Karolis Stulgys shipped for before this offer, across four continents
Work we take off your team
Five jobs, with the systems each one touches. None is a product on a shelf: we build the agent for the company that asks. Your work does not have to be on this list.
Describe a job that is not here- Brief to published post
A brief goes in, a drafted and edited page comes back, and publishing waits for a name.
Notion, your CMSHow a brief becomes a page - Draft and schedule
Your team asks in a channel. The agent drafts, queues and reports what the last batch did.
Slack, TypefullyHow a post gets drafted and queued - Support triage
Every new ticket gets read, grouped and answered where the answer is known, and filed as an issue where it is a bug.
Zendesk or Intercom, GitHubHow the triage works - Failed-payment follow-up
Overnight failures come back as a summary and a drafted message per customer. Sending needs a signature.
Stripe, your outboxHow the follow-up works - Store operations
Stock, pricing and order exceptions watched on a schedule, with the ones that need a decision brought to a person.
ShopifyHow the store is watched
What we shape your agent around
We take each of the five points below from your company. We write them down before we build, and your agent follows what we wrote.
Who asks
Which people can start your agent, from which channel, and what they may ask it for.
Who approves
Which actions stop and wait, and whose name goes on the approval when they do.
The systems it touches
The tools your team already runs. Your agent uses the permissions of the person who asked. There is no shared key.
Your words
The names your company uses for things. We take them from your issues, your docs and your tickets. The answers then read like your team wrote them.
What finished means
Your definition of a finished piece of work. We write it down, so you can hold your agent to it.
How we keep it right
The five points above are a promise. These five are how we prove it, every month, for as long as we operate your agent.
What an eval suite isThe eval suite
The written definition of correct for your agent. It runs on every change we make, and a failing suite stops the change. When we cannot describe correct for a task, we say so before we build it.
Approval gates
We approve changes inside your agent’s own repository and deployment. Every outward action waits for a named person at your company. That means anything reaching a customer, an inbox, a published page or a payment.
A named engineer
One of our engineers owns your agent and answers inside business hours. Your agent runs in your own cloud account, so its uptime is your platform’s. We do not promise uptime for software we do not run.
The spend cap
Every session has a spend cap that we agree with you. A session that reaches the cap and has someone to ask stops and asks. A session running unattended stops and reports.
The monthly report
What your agent did, what it cost, what we changed, and what we would do next. Cancel at the end of any month. Our work stops and your agent keeps running.
What your agent is built on
We build on Flue, the open agent harness the Astro organisation publishes. Your agent deploys into your own Cloudflare account, or onto Node.js where your platform needs it there instead. We resell nothing: the account, the keys and the code are yours.
Where your agent lives
A person addresses your agent by name in a chat channel. A system wakes it with an event. Both lists are what Flue verifies, so you can check them.
A person reaches it in
- Slack
- Microsoft Teams
- Discord
- Telegram
- Google Chat
- SMS through Twilio
Teams needs one step that Slack does not: an Azure Bot registration and an app password stored in your project. We do that work and quote it as its own line.
An event wakes it from
- GitHub
- Linear
- Notion
- Stripe
- Shopify
- Zendesk
- Intercom
- Resend
- Salesforce Marketing Cloud
Who puts their name on it
Karolis Stulgys founded the company in Vilnius. He owns the agents we build and signs the merges we make on a client’s repository.
The founder and six references“Karolis is the consummate professional. Highly personable, excellent communication skills, dedicated and technically astute. Along with all that he is a nice guy.”
Cathal McAliskeyLead IT Consultant, GemPoolWritten about his earlier work, before this offer existed. There is no case study for an agent we have built yet, and we would rather print nothing than invent one.
What AI agent development costs
A build price is a floor, because the work follows the number of systems your agent touches. Every monthly price buys the evals, the changes and the report.
Questions we get on the first call
The answers as we give them. Anything not here, ask on the call or in the brief.
You do. It lives in your repository and deploys into your own accounts, with your keys. Stop paying us and your agent keeps running. What stops is our work on it.
Every product prints its prerequisites next to its price, and each line says whether you already own it or we set it up. Nothing is hidden until the third day of a build.
Every session has a spend cap that we agree with you. A session that reaches the cap stops and asks you. A session running unattended stops and reports. The cap is our code, because the framework reports what a call cost and enforces nothing.
A named person at your company approves every outward action. That covers anything reaching a customer, an inbox, a published page or a payment. Our name never goes on something your customers see.
Flue, the open agent harness the Astro organisation publishes, deployed into your own Cloudflare account, or onto Node.js where your platform needs it. Four pieces are ours rather than the framework’s: the approval gates, the spend cap, the tool that opens a pull request, and the filter that decides what starts a run.
No, and nobody honestly can for software running in your own cloud account. We promise four things. A named engineer. An answer inside business hours. A green eval suite on every change we make. A monthly report of what your agent did.
We write an eval suite for every agent: the written definition of correct, run on every change we make. You can read it. Sometimes we cannot write down what correct means for something you want, and we tell you that before we build it.
No. One agent is one deployment with one name. A second name is a second agent, a second build and a second monthly line. One agent can still run several subagents behind its single name.
Yes. Teams needs one step that Slack does not: an Azure Bot registration and an app password stored in your project. We do that work and quote it as its own line.
You can, and the framework is open source. The framework is the easy part. The work is shaping the agent to your workflow, setting the approval policy, writing the eval suite, and holding all three steady while the framework moves under it.
Send us the issue you would hand to a new senior engineer
Or describe the work you want an agent to take over. We reply within one working day with a short spec and a fixed price.