I'm running a one-person AI consulting startup with Claude Code as my COO.
Not a metaphor — it actually runs operations.
Every morning, agent squads execute: research competitors, draft content, monitor costs, update memory.
I make decisions, Claude executes them across 16 domain squads.
What this actually looks like:
10 Claude Code sessions running in parallel right now
- 16 squads (marketing, engineering, finance, customer, etc.)
- ~100 agent definitions, all markdown files
- Autonomy and delegation with Task (subagents)
- Smart organization of agents, skills and permissions
- Hooks usage to local store and track the Goals, kpi, task resolutions, token economics.
- Shared memory via Postgres, session coordination via Redis
- squads dash shows what's running, what it cost, what changed
- Claude is aware of the operation and priorities
Why parallel sessions need infrastructure?:
One Claude session is easy. But when you have 10 or more running — one researching, one writing, one monitoring costs — they need shared state.
Otherwise they overwrite each other or duplicate work.
Agents = Markdown files (.agents/squads/*.md)
Memory = Postgres (persists across sessions)
Sessions = Redis (coordination, locks)
Costs = OpenTelemetry (what each run cost)
The results:
My GitHub contributions are up 10x since switching to this setup. Not because I'm working harder — because the COO handles the grunt work while I focus on decisions.
Why we built it this way:
We're dogfooding. If we're going to sell AI agent implementations to clients, we should run on them ourselves.
Every pain point we hit, every failure mode — that's consulting IP.
Honest status:
- 16 squads running daily operations
- 6 parallel Claude sessions, fully coordinated
- 10x GitHub contributions
- Every agent run tracked and costed
- First consulting clients in pipeline
This is what running a company with an AI COO actually looks like in January 2026.
check all what we were able to build since Opus 4.5 (11-24-25), the best COO llm so far.
It is a metaphor, though. If you end up running afoul of any laws or other regulations, get sued, fined, etc, Claude is not a person who is accountable for those consequences - you are.
I thought into that too! but not yet.
as you want the agents perform bash commands to interact with differnts services, by now you need to run it on your local computer and terminal where is assumed you'll handle properly all secrets.
but, yes! I imagine too in a near feature talking with my headphone to a my local computer and coordinating from there when I'm afk
but they are supporting lot of our task, researching, intelligence reports, content writing, website and cli improvements, among others.
I'll say the key to trust and being able to delegate is the alignment of the agents with your goals.
This post, explain the difference between the 2025 data agents (only data retrievers) ; and what' possible today (agents that act based on trusted data)
Every morning, agent squads execute: research competitors, draft content, monitor costs, update memory. I make decisions, Claude executes them across 16 domain squads.
What this actually looks like:
10 Claude Code sessions running in parallel right now - 16 squads (marketing, engineering, finance, customer, etc.) - ~100 agent definitions, all markdown files - Autonomy and delegation with Task (subagents) - Smart organization of agents, skills and permissions - Hooks usage to local store and track the Goals, kpi, task resolutions, token economics. - Shared memory via Postgres, session coordination via Redis - squads dash shows what's running, what it cost, what changed - Claude is aware of the operation and priorities
Why parallel sessions need infrastructure?:
One Claude session is easy. But when you have 10 or more running — one researching, one writing, one monitoring costs — they need shared state. Otherwise they overwrite each other or duplicate work.
Agents = Markdown files (.agents/squads/*.md) Memory = Postgres (persists across sessions) Sessions = Redis (coordination, locks) Costs = OpenTelemetry (what each run cost)
My GitHub contributions are up 10x since switching to this setup. Not because I'm working harder — because the COO handles the grunt work while I focus on decisions.Why we built it this way:
We're dogfooding. If we're going to sell AI agent implementations to clients, we should run on them ourselves. Every pain point we hit, every failure mode — that's consulting IP.
This is what running a company with an AI COO actually looks like in January 2026. check all what we were able to build since Opus 4.5 (11-24-25), the best COO llm so far. Happy to answer questions about what works, what doesn't, and what surprised us.reply