Case study · Multi-agent system · Owlab

How we run our own company with a team of AI agents

Owlab is 15 people, and for years the founder held sales, follow-ups, payments and client care in his own head. So we built ourselves what we build for clients: a team of AI agents on the Claude Agent SDK, each owning one part of the business. They watch, remind and prepare. A person decides and sends.

All case studies
Illustration, AI-generated.
Client
Owlab itself
Service
AI agent development, multi-agent system
Built on
Claude Agent SDK (Anthropic), Claude models, in-process MCP tools, Python, Telegram
Connected systems
Pipedrive, Gmail, Google Calendar, Jira, Slack (read only), Fireflies, our own dashboards
Status
in daily use; figures below cover 1 October to 12 November 2026

The situation

In a small agency the founder is the bottleneck. Every deal needs a follow-up, every invoice a reminder, every client a check-in, every project a look at what has stalled.

None of it is hard. All of it is easy to forget on a busy week. And forgetting one follow-up can cost a deal.

We did not want another dashboard to check. We wanted colleagues that come to us.

Colleagues that come to you, not another dashboard to check.
Colleagues that come to you, not another dashboard to check. Illustration, AI-generated.

What we built

Five AI agents, each in its own chat, each with one lane.

  1. 01Head of Sales

    Watches the whole pipeline in Pipedrive, plans follow-ups with new content each time, looks for upsell with existing clients.

  2. 02Customer Success

    Looks after client satisfaction, asks the right question at the right time and keeps a relationship log. It does not sell: when it sees money, it hands over to sales.

  3. 03Finance

    Keeps receivables and cash in view, chases unpaid invoices on a fixed ladder and flags what was sold but not yet invoiced.

  4. 04Operations

    Watches delivery: stalled tasks, estimates that run over, missing worklogs. It reads Jira, Slack and meeting notes, and writes nowhere.

  5. 05Reviews and directories

    Keeps the rhythm of client reviews and our profiles in business directories.

Each agent wakes up on its own schedule, reads its systems and speaks only when there is something worth saying.

Five lanes: Head of Sales for pipeline and follow-ups, Customer Success for satisfaction, Finance for receivables and cash, Operations for delivery, and the founder who reads, approves and sends. None of the agents can message a client.
Each agent reads its own systems, such as the sales pipeline.
Each agent reads its own systems, such as the sales pipeline. Illustration, AI-generated.

A person stays in charge

None of these five agents has a tool that sends a message to a client. They prepare the text, a person reads it and sends it.

Each agent sees only its own tools. Finance data is hidden from every other agent, and a lane lock in code stops one agent from closing another agent's tasks.

Every tool call passes a permission gate first. In six weeks the gate blocked 89 calls that were not allowed.

Coding sessions and agents work as one team

When a decision is made in a coding session, for example a client email sent or a price agreed, it is written to the agent's handoff file. The agent wakes up, reads it and carries on from there.

That happened 174 times in six weeks. Nobody had to retell anything.

What broke, and how we fixed it

  • An SDK update silently stopped calling our permission check. We moved the gate into a pre-tool hook and proved it on the live SDK.
  • A restart could cut an agent off mid-answer. Now the current answer finishes before the service stops.
  • In long sessions an agent dated its notes by the day the chat started. Now the date and time travel with every message.
  • A deploy with a merge conflict broke the CRM tools for every agent at once. Now every deploy runs the tests on the server and restarts only on green.
  • Agents drew financial charts by hand and sometimes got a number wrong. Now code draws every chart from the data, and the agent never retypes a figure.

Results

From 1 October to 12 November 2026:

  • The agents woke up on their own 2,985 times and wrote to the founder 1,486 times. The rest of the time they had nothing worth saying, and said nothing.
  • They created 249 follow-up tasks in Pipedrive.
  • 174 handoffs from coding sessions reached the right agent without a word retold.
  • The permission gate blocked 89 tool calls.

Under the hood: 85 tools across 15 in-process MCP servers, and 1,687 automated tests, 195 of them about safety.

From 1 October to 12 November 2026: agents woke up 2,985 times, sent 1,486 messages to the founder, created 249 follow-up tasks, received 174 handoffs from coding sessions, and the gate blocked 89 tool calls.
The team works from what the agents surface.
The team works from what the agents surface. Illustration, AI-generated.

Want a team of AI agents around your own CRM, finance or delivery?

We start with a fixed-price pilot from $1,000: one agent, one lane, on your real data.

Book a call about your process
Your process

Want a team of AI agents around your own CRM, finance or delivery?

Tell us in two lines. We'll reply with how an agent would take it over and what the first step would cost.

  • A fixed-price first step, from $1,000
  • Half at the start, half after we check it on your data
  • NDA on request
  • [email protected]Vitaliy, CEO
  • Vesivärava tn 50-201, 10152 Tallinn, EstoniaHead and development office

By sending, you agree to our Privacy Policy.