How I Built
A C-suite Agent Swarm OS
in Five Days
A Portfolio Executive's walkthrough of building a 13-node executive operating system. Twelve C-suite supervisors plus a Chief of Staff and a CEO integrator. Each vertical is a supervisor-swarm-auditor unit. Each produces one schema-locked JSON report per day. The CEO sees one screen, makes three calls.
Yuri Kruman
Portfolio executive, 3x CHRO · Jun 2026
supervisor nodes
executor agents under them
CEO calls a day, by design
to all 13 nodes operating
The 30-Second Version
The C-suite Agent Swarm OS is a full executive operating system. Twelve C-suite supervisor agents (CMO, CRO, CFO, CPO, CTO, COO, CSO, CCO, CLO, CDO, CHRO, Chief of Staff) plus a CEO integrator. Each vertical follows the supervisor-swarm-auditor pattern: executors do the work, an auditor scores the output, the supervisor synthesizes. All 13 supervisors produce schema-locked JSON reports. The CEO node integrates the 12 specialist reports into 3 daily calls and a weekly thesis update. The whole system is Shabbat-aware.
The Problem
Solo operators of 8+ properties unable to keep an executive function's worth of attention on each.
The Stack
Claude Code subagents + Node + JSON schemas + Windows Task Scheduler.
What It Doesn't Need
No agent framework. No vector DB. No agent platform. No paid team.
Build Time
Five days: 1 for the schema and pattern, 1 per cluster of three verticals, 1 to wire the CEO integrator.
If you operate a portfolio of three-plus properties and have ever felt that none of your "functions" got the brain time they deserved this week, this build log is for you. The pattern compounds. Each new vertical is a copy-paste of the supervisor-swarm-auditor unit with a new system prompt.
Why an Executive OS, Not Smarter Spreadsheets
I run eight properties simultaneously. AI HR Pilot, DueDrill, BookToCourse.AI, the Beast Score / Leverage Brief stack, Ohr Vishua, the 92% Fund, the PE-fund-website business and consulting. Plus a Fortune-500 fractional CHRO seat. Plus the book and podcast. By Q2 2026 I had hit the wall of human attention. The number of functions a portfolio operator should track exceeded the number of functions I could think about in one morning.
12
C-suite functions a normal company staffs with humans
1
human (me) running 8 properties
~3
of those functions getting real attention each week
9
silently decaying because of inattention
The fix was not "work harder." The fix was an agent swarm that staffs each function with a daily-firing supervisor. The supervisors are not full executives; they're specialist attention. They watch their function 24/7, produce one schema-locked report at 07:00 every weekday and let me know when something needs the CEO's call. Most days, most functions are GREEN and I never think about them.
- 1One vertical per C-suite function. CMO, CRO, CFO, CPO, CTO, COO, CSO, CCO, CLO, CDO, CHRO + Chief of Staff.
- 2One CEO node that integrates. Reads 12 reports. Produces 3 calls a day. Maintains a weekly thesis.
- 3Effectiveness over efficiency. A day is successful only if dollars moved, risk dropped or strategy clarified. Encoded in the CEO prompt.
Every vertical implements the SAME supervisor-swarm-auditor pattern. No exceptions.
CFO might feel different from CMO. It is not. Same pattern: supervisor, executors, auditor, schema-locked JSON. The pattern is the moat. The pattern is what makes the CEO integrator possible. If you let one vertical innovate on the pattern, the integrator breaks within a week.
The Stack (and Why Each Piece)
Click each layer.
Contract
roles/_shared/output-schema.json
The single output contract every supervisor satisfies. Fields: role, date, status RAG, top_movements, ceo_asks, conflicts, thesis_alignment. The contract is what makes the CEO integrator possible.
Pattern
roles/_shared/PATTERN.md
The universal supervisor-swarm-auditor spec. Every vertical's agents/, workflows/, contexts/, RUN.md structure is identical. New vertical = copy the folder, change the prompts.
Verticals
12 specialists + Chief of Staff
CMO (marketing), CRO (revenue), CFO (cash + runway), CPO (product), CTO (build + uptime), COO (vendors + tools), CSO (market intel + portfolio), CCO (retention + churn), CLO (legal + IP + SEC), CDO (KPI + attribution), CHRO (solo-founder time + boundaries), Chief of Staff (standup runner + brief author).
Integrator
CEO node (Opus)
Reads the 12 specialist JSON reports. Produces 3 calls today, arbitrations and a weekly thesis. Documented in detail in the CEO Command Center build log.
Cadence
Sun-Thu 07:00 IDT, Shabbat OFF
07:00 the 12 verticals fan out. 07:30 Chief of Staff synthesizes. 07:45 CEO ingests. 08:00 brief lands. Fri 14:00 pre-Shabbat wrap. Sat OFF. Sat 21:00 (post-Shabbat) Sunday queue lock.
Decision Log
Append-only JSONL
Every CEO call is one JSONL line: ts, call, rationale, specialist_inputs, expected_outcome, review_date. Never edited. Weekly retrospective reads this file.
The Six-Phase Build Sequence
Each phase is a focused day. Order matters: schema before pattern; pattern before any vertical; first cluster before any new variant.
Schema + Pattern (the constitutional layer)
Write roles/_shared/output-schema.json and roles/_shared/PATTERN.md. Every later vertical conforms to both. The schema is enforced by auditors. The pattern is enforced by the folder layout.
What I'd Do Differently Today
Treat CHRO output as a hard constraint, not an FYI
CHRO is the function that catches "Yuri is overworked and the system is about to fail." I treated its output as informational for the first three weeks. Now its RED status hard-blocks the CEO from adding new asks.
Build CFO before CMO
I started with CMO because marketing is fun. CFO would have saved me 3 days of CMO suggesting things I couldn't afford.
Per-vertical context size budget
CSO loves to stuff its prompt with market intel until the cost spikes. Set a max-tokens-per-vertical-per-day budget; auditor enforces.
Ship Cluster 1 in production before authoring Cluster 2
Authoring all 12 verticals before running any of them is the temptation. Resist. Cluster 1 in production reveals what Cluster 2 actually needs.
Adapt This for YOUR Org Size
13 nodes fits a portfolio executive. Smaller orgs run smaller swarms. The pattern is the same; the number of verticals changes.
| Persona | Verticals | Integrator |
|---|---|---|
| Solo creator | Content, audience, ops (3) | Self (no CEO node needed) |
| Pre-seed founder | Product, growth, fundraising, ops (4) | CEO node, weekly thesis |
| Seed/Series A | CMO, CRO, CPO, CTO, CFO, CHRO (6) | CEO + CoS |
| PE / VC partner | Deal flow, IC, port co updates, LP (4) | Managing partner integrator |
| Portfolio executive | 12 + CoS | CEO node, daily 3 calls, weekly thesis |
Starter Prompts for Claude / Cursor
Four prompts to take a swarm OS from blank repo to running this week.
"Write roles/_shared/output-schema.json (JSON Schema draft-07) for an executive-OS specialist report: role, date, status RAG, top_movements array (label/delta/confidence), ceo_asks array (ask/dollar_impact_if_ignored/deadline), conflicts array (other-role names), thesis_alignment 0-1. Then write roles/_shared/PATTERN.md describing the supervisor-swarm-auditor pattern with the per-role folder layout: agents/, workflows/, contexts/, RUN.md. Include 3 PASS examples and 3 FAIL examples for the schema."
"Write scripts/scaffold-vertical.mjs that takes a role name (e.g. CFO) and scaffolds roles/cfo/ with: agents/supervisor.md (placeholder), agents/executor-*.md (3-5 placeholders), agents/auditor.md (placeholder), workflows/daily.js (fan-out + audit + supervise), contexts/brand-voice.md (symlink), contexts/role-specific.md (placeholder), RUN.md. The scaffolder reads roles/_shared/PATTERN.md and ensures the new folder conforms exactly."
"Write workflows/daily-executive-standup.js (ESM Node). 07:00: fire all 12 vertical workflows in parallel; each runs its own supervisor-swarm-auditor pipeline; each writes data/reports/YYYY-MM-DD/
"Modify the CEO integrator prompt so that if today's CHRO report.status === 'RED', the integrator MAY NOT add new CEO asks beyond the existing decision log. It may only resolve open items, drop scope or push deadlines. The integrator must explicitly reference the CHRO RED in the day's brief and propose a recovery plan that reduces total load by ≥ 30% within 5 days."
What the C-suite Swarm OS Is Not
It is not a replacement for human executives. It is not "AI cofounders." It is not autonomous; every executor produces drafts, every supervisor reports, every CEO call is approved by the human CEO before any external action.
What it is: an attention amplifier. It keeps 12 functions in working memory at once so a solo portfolio operator can run what would normally take a real C-suite. The narrowness is the point. Systems that try to also EXECUTE blow up. Systems that REPORT to a human integrator compound.
The question is not
"can AI run a company?"
The question is:
"What functions in my portfolio are silently decaying because no one (including me) watches them daily?"
List those functions. Stand up a supervisor-swarm-auditor unit for each. One JSON report a day. One CEO node to integrate. Five days from blank repo to running.
This walkthrough is part of the Portfolio Leverage Co. Build Bench series. For the cohort where we build these together, apply here.