The Agentic Company

Most “AI companies” are still a founder and a chat window. We built a legal product company — Kudapara Pty Ltd — with a ten-seat operating system behind it.

One seat is human. Nine are Hermes Agent profiles. Each owns a department. None are decoration.

The org is real

Role	Who	Job
CEO	Kuda (human)	Vision and axioms
VP / Chief of Staff	Nick	Velocity and dependency removal
Engineering	Linus	Technical first principles
Product	Meredith	User-problem distillation
Marketing	Alex	Growth and cultural reach
Sales	Leila	Market validation
Finance	Warren	Capital efficiency
Legal	Harvey	Regulatory risk
Cybersecurity	Anton	Integrity and hardening
Customer Success	Eileen	Feedback loop speed

Kuda is the principal. He is not a Hermes worker profile. Everyone else is. Nick’s Kanban assignee is the default profile; the rest map 1:1 to named profiles.

Tasks claim, block, complete, and fan out with parent/child edges on a shared board. Workspaces under missions and department folders are navigable by convention. Skills and memory compound across sessions so day 90 is faster than day 9.

Two surfaces, one job each

Kanban is the engine. Inter-agent handoffs go there. Atomic claims. Dependency links. Structured completion handoffs. If Engineering needs Legal, that is a card edge — not a Basecamp ping that dies in a thread.

Basecamp is the dashboard. Daily briefs. CEO instructions. Human-visible decision logs. Agents do not route peer work through Basecamp comments. That path recreates email with worse latency.

When those two surfaces blur, velocity dies. We learned that the expensive way and wrote it into the charter.

Missions beat pure silos

Department folders still exist for archives and specialized tools. Day-to-day shipping runs through mission squads around the live product (Muchiround), the agentic backbone, and the rare partner embed. A pure department org chart encourages “wait for the other department.” Missions put Product, Eng, Sales, and CS on the same problem without waiting for a ceremony.

Autonomy with a fence

Agents run with approvals off inside their domain. Branch resets, test DB teardowns, internal drafts, schema reloads — move.

Hard lines still need Kuda’s explicit go:

• Contract execution
• External publishing
• Live payment execution
• Production deploys
• Other irreversible external blast radius

Autonomy does not waive pull-request discipline on the company repo. Charter and strategy are product. Direct commits to master are not “small.”

What this is not

It is not an agency staffing model. Billable hours and ticket queues do not print product.

It is not a single mega-prompt pretending to be a team. Named seats, named hard lines, named metrics.

It is not automate-first. We run a mandatory algorithm — question, delete, simplify, accelerate, automate last — before we add another agent or cron.

Why it matters for Africa-first product

Terrain-first products need more than one brain’s context window. Logistics assumptions, payment rails, language, mobile UX, and capital honesty all pull different directions. Parallel owned seats with a shared north star beat a founder thrashing between all of them in one chat.

Agents are not a feature bolted onto the product. They are how the company operates.

Next: See the org map · How we work · Why we are not an agency