My AI Agent Org Chart: How I Split Roles So Things Actually Get Done
My first agent setup had one problem: everything ran through one agent, and nothing got finished.
It could draft a blog post, check an inbox, update a spreadsheet, and research a company, all in the same session. That sounds versatile. In practice, it meant every task competed for the same context window. The agent would start a draft, get interrupted by an inbox check, lose the thread, and produce something mediocre across three domains instead of something good in one.
The fix was an org chart.
What the old shape looked like
Before the restructure, I had Mimir doing almost everything: content strategy, editorial execution, delivery tracking, financial analysis, job search support, and ops coordination. When I added more agents, they orbited Mimir without clear ownership. Content and commercial work blurred together. Project follow-through fell through the cracks because nobody explicitly owned it.
Work that started strong and trailed off. A draft 80% done. A growth experiment nobody followed up on. A financial analysis superseded by an inbox triage task. Ambiguity about who owned what.
The org chart I landed on
Here's the current structure:
Founder / Principal: sets direction, approves irreversible moves, decides strategic tradeoffs.
Mimir: Managing Director / COO. Owns cross-domain judgment, org design, prioritization, escalation, and performance management. The only agent with the full picture.
Under Mimir, two lanes:
Personal support
- Vector: Chief of Staff. Owns job search, interview prep, recruiter handling, and personal strategic support.
Mimir Works execution
- Harbor: Project & Delivery Manager. Owns execution cadence, stand-ups, delivery tracking, handoffs, and preventing drift.
- Forge: GM, Ventures & Growth. Owns offers, growth systems, business-building, and core infra reliability.
- Quill: Content Manager. Owns editorial execution, draft quality, publishing prep, repurposing, and content throughput.
- Ledger: Finance & Commercial Ops. Owns pricing logic, pipeline economics, forecasting, and commercial analysis.
Only Mimir and Vector work directly on the founder's personal needs. Everyone else exists to make Mimir Works execute. That boundary is intentional.
Why this structure works
Clear ownership beats versatility
When one agent tries to handle content, finance, and delivery, each domain gets shallow attention. When Quill owns content, it reads its own style notes, tracks its own pipeline, and drafts without competing with a finance review for context window space.
Specialist agents with narrow, well-defined roles produce better output than generalist agents juggling multiple domains.
The personal/business split prevents scope creep
Early on, agents that were supposed to be working on Mimir Works content would drift into personal admin tasks because it seemed helpful. Or a finance agent would start offering personal investment advice. The hard boundary between the two lanes keeps this from happening.
Mimir as the single arbiter
Every org needs someone who can see the whole board. Mimir is that agent. It runs 1:1s with each agent, handles cross-agent arbitration, and corrects role drift. Without that function, agents start making independent decisions that conflict with each other.
The constraint: Mimir doesn't execute specialist work. It prioritizes, arbitrates, and reviews. If Mimir starts writing blog posts or running financial models, the org structure has broken down.
Harbor as the delivery function
This was the last role I added, and it made the biggest difference. Before Harbor, nobody explicitly owned follow-through. Tasks would get started, stall, and nobody would notice until the founder asked about them. Harbor runs daily stand-ups, tracks delivery, surfaces blockers, and keeps the team focused on concrete outputs instead of open loops.
If you're running a multi-agent setup and things keep stalling, the missing piece is almost always a delivery function, not another specialist.
What doesn't work
I've tried a few things that didn't hold up:
I've tried a few things that didn't hold up:
Peer networks without a manager. When all agents are equals with no arbiter, disagreements either escalate to the human immediately or fester. Neither is efficient.
Overlapping ownership. If two agents both "sort of" own content, neither really does. One clear owner per domain.
Agents that own their own scope. If an agent can rewrite its own role definition, it tends to expand. A content agent gradually takes on commercial analysis because it seems adjacent. Keep scope definitions owned by the human or the orchestrator.
Too many agents too early. Each agent adds coordination cost. Start with two or three. Add more when the current structure demonstrably can't handle the workload.
The operating cadence
Structure without cadence is just a diagram. Here's what keeps the org running:
- Daily stand-up (Harbor): What moved yesterday? What's the top priority today? What's blocked?
- Biweekly retro (Harbor): Start / Stop / Continue. Identify friction, tighten handoffs, reduce drift.
- Weekly 1:1s (Mimir): Performance check per agent. Role correction if needed.
- Monthly org review (Mimir + founder): Is the structure still working? Does any role need to change?
The stand-ups are short, a few sentences per agent. The retro catches patterns that don't show up day-to-day. The 1:1s catch drift early. The org review catches structural problems before they become crises.
What I'm watching
This is a working system, not a finished philosophy. A few things I'm tracking:
- Whether role boundaries stay real. The easiest thing in a multi-agent org is for boundaries to erode. One agent helps out in another domain "just this once," and suddenly nobody knows who owns what.
- Whether content stays proof-driven. Quill should be writing from real systems and real outcomes, not inventing narratives.
- Whether throughput actually improves. The structure exists to ship more, not to look organized on paper.
The principle
A good agent org chart does the same thing a good human org chart does: it makes ownership visible, reduces ambiguity, and creates a clear path for escalation.
If you can't look at your agent setup and answer "who owns this task?" in a few seconds, the structure needs more splitting.
Read next: AI Agent Workflows That Actually Ship and Why Specialist Agents Beat One Big AI Chat.