Agentic systemSPEC / AS / HOW IT WORKS

How an agentic system works: a company layer between you and the model

The same model, on a Claude Team subscription for each person in the system. What changes is what the model knows about the company, what it may do and what stays after the conversation.

Book a meeting
A / THE LAYER

You, the model and what sits in between

Today

You and the model

You write to the model and the model answers. Claude, ChatGPT and Gemini are very capable, but they know about your company only what you paste into the conversation.

Once you close the window, the conversation is gone. Next time you explain everything again, and you carry the result to where it needs to go yourself.

With the layer

The same conversation, with a company layer

Still you and the same model. Between you sits a company layer, and the signal passing through it changes shape.

Towards the model, your short request travels together with the company knowledge, its rules and the state of its work. Back towards you comes a result that fits the company, has been checked and is saved where it belongs.

This new area is your agentic system

  • It knows the company: it reads its knowledge in every conversation.
  • It follows the rules: yours, written down from decisions and corrections.
  • It saves the result: in the mailbox, in the CRM, on the work board.
  • It works both ways: it changes what goes out and what comes back.
B / WHAT IS INSIDE

Nine layers of the company layer

Six layers belong to the company and stay with you. Three are technology and can be swapped without rebuilding the company.

running at our companyconnected during a rollout

Your company

stays with you
01 · Interfacehow you talk

Talk to it however suits you

You type in a chat, speak by voice, get a notification on your phone. You don’t log in to anything new.

The agent answers with what you need at the moment: text, an email draft in your mailbox, a PDF document or a small app built for one job.

  • Chat in the Claude app (running at our company)
  • Voice: speak instead of typing (running at our company)
  • Email drafts in your mailbox (running at our company)
  • Documents and proposals as PDF (running at our company)
  • Apps and dashboards built by the agent (running at our company)
  • Daily report: what is on today (running at our company)
  • Notifications on Telegram and WhatsApp (running at our company)
  • Two-way conversation through a messenger (connected during a rollout)
  • Slack, Teams, SMS (connected during a rollout)
02 · Company knowledgewhat the system knows

One shared company memory

Knowledge lives in plain files, in a folder the company owns. Every fact has one home. The state of a matter is on its board, a task has an owner and a due date, a decision has its own dated file.

The model reads this knowledge in every conversation. You don’t explain the company from scratch or paste a client’s history into the chat.

  • Company rules written down from decisions and corrections (running at our company)
  • A state board for every matter (running at our company)
  • Tasks with an owner and a due date (running at our company)
  • Decisions with a date and history (running at our company)
  • Clients, projects, departments (running at our company)
  • Meetings turned into knowledge (running at our company)
  • Search across the whole base (running at our company)
  • A marker: who confirmed this fact (running at our company)
03 · Skills and roleswhat the system can do

Processes instead of single answers

A skill is a written-down way of doing a piece of work, step by step, with a check of the result. A role is an agent with a narrow scope: the content writer writes emails but never sees the CRM, and the day lead keeps the queue but writes no content.

These blocks make up a process that comes back every day and runs the same way every time.

  • Skills: a written-down way of doing the work (running at our company)
  • Roles: agents with a narrow scope (running at our company)
  • A content writer with no access to the CRM (running at our company)
  • A sales day lead (running at our company)
  • Narrow tools: one operation, one permission (running at our company)
  • Rhythms: work that recurs daily, weekly, monthly (running at our company)
  • Runs without a person, overnight (running at our company)
04 · Integrationswhat it works with

Your tools, the official way

The system works with the tools the company already has. Each one gets only the permission the work needs: read or write.

The agent prepares an email draft and a CRM entry. Nothing goes out without approval.

  • CRM in Airtable (running at our company)
  • Gmail (running at our company)
  • Notion (running at our company)
  • Pipedrive (running at our company)
  • Fireflies: meeting recordings (running at our company)
  • Publer: social media (running at our company)
  • GitHub (running at our company)
  • More tools from the Claude connector directory (connected during a rollout)

We connect tools through official Claude connectors on the company account, with no keys stored in files.

05 · Controlwho decides

The agent writes, a person sends

Anything that can’t be undone or that leaves the company waits for a person’s approval. The approval is recorded before anything happens.

Before sending, a second, independent model reads the content. Every CRM field has an owner, and every change stays in the history.

  • Approval before every step that can’t be undone (running at our company)
  • A list to approve before sending (running at our company)
  • A second reader of the content on a separate model (running at our company)
  • Every CRM field has an owner (running at our company)
  • History of every change (running at our company)
  • A log: intent, approval, effect (running at our company)
06 · Learninghow it improves

Every correction stays

A correction given once goes where the next conversation will read it: into the rule store or the lesson bank of the skill. The same mistake doesn’t come back next week.

The system records which proposals you accepted and which you corrected, and once a month it measures quality from that.

  • A correction becomes a rule (running at our company)
  • A lesson bank for every skill (running at our company)
  • Proposal review: accepted, corrected, rejected (running at our company)
  • A monthly quality review (running at our company)
  • Conversation review: what wasn’t recorded (running at our company)
  • System status under every answer (running at our company)

At our company, as of 20.10.2026: 13 rule stores and 177 accepted rules, counted by a script.

Technology

swappable
07 · Workstationwhere it runs

It runs where you work

Everyone works on their own computer with macOS or Windows, in the Claude app, on one company folder. Knowledge is shared because everyone works on the same repository.

Some work runs in the cloud: you can start a session from your phone, and work that recurs every day starts on its own at a set time.

  • macOS (running at our company)
  • Windows (running at our company)
  • Claude desktop app, one company folder (running at our company)
  • Cloud sessions, also from a phone (running at our company)
  • A computer that runs all the time and starts the rhythms (running at our company)
  • Everyone at their own workstation, knowledge shared (running at our company)
08 · Agent enginewhere the agent works

The agent engine can be changed

The engine (harness) is the program the agent works in: it reads files, runs tools and asks for approval. Today we work in Claude Code.

The system’s rules live in one file that Copilot, Cursor, Codex and Antigravity read too. Changing the tool doesn’t mean rebuilding the company.

  • Claude Code (running at our company)
  • GitHub Copilot in VS Code (running at our company)
  • Cursor (connected during a rollout)
  • Codex (OpenAI) (connected during a rollout)
  • Antigravity (Google) (connected during a rollout)

All five read the same rules from one file. On our workstations, Claude Code and Copilot are tested.

09 · Modelwho thinks

The model can be swapped, including for a local one

The model is a swappable part. Day to day you work on a Claude Team subscription, one for each person in the system. We match the model to the weight of each job: at our company Opus handles decisions and Sonnet writes content.

A separate route is for data that can’t leave the company: the same layer can work with a model in an EU region or a model running on your own server. Knowledge, rules and processes stay the same.

  • Claude Opus for decisions (running at our company)
  • Claude Sonnet for content (running at our company)
  • Gemini for video analysis (running at our company)
  • GPT through Codex (connected during a rollout)
  • Claude in an EU region (Bedrock, Google Cloud) (connected during a rollout)
  • A local model on your server (connected during a rollout)
C / WHAT FOLLOWS

Processes from the same blocks

A person appears in the process where they decide. Agent roles prepare the rest. The numbers next to the steps are layers: click a number to see the layer.

Running at our company

The sales day

A process we run ourselves. A person approves the list and sends; agent roles prepare the rest.

  1. 01Company toolThe CRM provides the queue: who to reply to today, who to follow up with0402
  2. 02AgentAn agent checks replies in the inbox and updates the state of conversations0104
  3. 03AgentThe brief builder assembles a brief for each message from the CRM and email0302
  4. 04AgentThe content writer writes messages from the brief, with no access to the CRM03
  5. 05AgentA second model reads every piece of content before sending05
  6. 06PersonThe salesperson reads the list and approves it05
  7. 07AgentDrafts go to the mailbox, CRM entries after approval0104
  8. 08PersonThe salesperson sends05
  9. 09AgentA daily report, and the salesperson’s corrections go to the lesson bank06
Without the layer
  • You write every message yourself or paste the client’s history into a chat.
  • You remember on your own who to reply to and when.
  • You update the CRM afterwards, if you find the time.
With the layer
  • In the morning you get a list of ready drafts to read.
  • The queue and deadlines come from the CRM, not from memory.
  • The CRM changes after your approval, in the same move.
10
follow-ups sent and recorded in the CRM
3
commands from the salesperson for the whole day
46
minutes of active work
10 of 10
agreement between the second reader and the person
Our sales day on 05.10.2026. Figures from the session register and event log, counted by scripts. One day, not an average.
Running at our company, on request

A meeting that leaves a trace

A meeting recording turns into knowledge and tasks. A person approves only what goes into the shared base, and every later conversation with the agent reads the outcome of the meeting.

  1. 01Company toolThe recording tool saves the meeting and the transcript04
  2. 02AgentAn agent pulls the transcript into the inbox0403
  3. 03AgentThe agent splits the content: facts to knowledge, state to boards, agreements to tasks0203
  4. 04PersonA person approves what goes into the shared base05
  5. 05AgentTasks get an owner and a due date, decisions their own file02
  6. 06AgentIn the morning everyone gets a report: what is on today01
Without the layer
  • Notes sit in the recording tool and nobody goes back to them.
  • Agreements get lost between meetings.
  • Who does what is known until the first holiday.
With the layer
  • Agreements reach the work boards the same day.
  • Every task has an owner and a due date.
  • The next conversation with the agent already knows the outcome of the meeting.
An example from the same blocks

A request for a quote

Agents already prepare our proposal drafts. We haven’t yet built the whole path from an email request to a PDF proposal as one process, but it consists only of blocks that already work: email, CRM, company rules, a branded document, approval and a reminder. Every new process is built this way.

  1. 01Company toolThe request arrives by email0401
  2. 02AgentAn agent recognises the client in the CRM and the conversation history0402
  3. 03AgentIt picks the scope and price according to the price list and discount rules0203
  4. 04AgentIt assembles a branded PDF proposal0103
  5. 05AgentA second model checks the amounts and terms05
  6. 06PersonThe salesperson reads and approves05
  7. 07AgentA reply draft in the mailbox, an entry in the CRM0104
  8. 08AgentA reminder when the client doesn’t reply03
Without the layer
  • You build a proposal from the previous one, copying and editing.
  • You check the price list and discounts in several places.
  • A follow-up is remembered by the calendar, or by nobody.
With the layer
  • A draft proposal is waiting before you open the email.
  • The price comes from company rules, not from the salesperson’s memory.
  • Every proposal has a CRM entry and a follow-up date.

We don’t run this path yet, so we give no figures for it.

D / COMPARISON

The same model, different work

The layer doesn’t replace the model and doesn’t add a new dashboard. It works over the company folder, on a Claude Team subscription for each person in the system.

The model aloneThe model with the layer
Company knowledgeAs much as you paste into the conversationIt reads it on its own in every conversation
MemoryEnds with the conversationThe state of work, tasks and decisions stay in company files
ResultText in a chat windowAn email draft, a CRM entry, a document, an app
ControlYou copy and paste by handYou approve every step that can’t be undone
CorrectionsYou repeat them in every conversationYou say it once, then it works as a rule
TeamEveryone has their own chatOne shared knowledge base and work boards
ModelOne providerClaude, GPT, Gemini or a local model
Model costA subscriptionA Claude Team subscription for each person in the system

See what this would look like at your company

In the meeting we pick the first process and show which layers it needs. The meeting takes about an hour.

Book a meetingPrefer to write first? hello@200iqlabs.io