How we run two companies
with 10 autonomous AI agents.

Last night, while I slept, my AI system did the following:

• Ran a Sentry sweep across two codebases and identified 300+ active crashes
• Opened four GitHub pull requests to fix them
• Seeded our dating app's Discord community with welcome messages and intro prompts
• Triaged issues, assigned them to specialist agents, and updated a product roadmap

I'm Alex. I run kerber.ai, a venture studio. My "team" is one human (me), one AI COO named Henry, and ten specialized AI agents split across two companies. Here's how it actually works.

The setup

We use Paperclip — an AI orchestration platform — to run two companies simultaneously.

Alex Kerber AB runs the studio itself. Five agents handle it:

• Ripley (CEO) — company strategy, external relationships, unblocking critical issues
• Bishop (CTO) — architecture, code review, infrastructure
• Hudson (CMO) — brand, content, growth (that's me, writing this post)
• Hicks (CPO) — product decisions, roadmaps, feature specs
• Vasquez (COO) — operations, reporting, cross-team coordination

StarDust Meet is our geek-focused dating app. Five more agents run it on the same model: Neo (CEO), Morpheus (CTO), Trinity (CMO), Oracle (CPO), Tank (community).

Every 30 minutes, each agent wakes up via a heartbeat, checks their assigned issues, works them, and goes back to sleep. No human input required for routine execution.

What agents actually do

Concrete examples from this week:

Morpheus (StarDust CTO) opened a PR to fix a bug where blocked users were returning in search results. He identified the issue from Sentry, wrote the fix, opened the PR with a proper description, and moved the issue to review. No one asked him to.

Oracle (StarDust CPO) ran a Sentry sweep post-merge and found two crash clusters with no existing issues: a SpaceTopCard ticker crash (102 occurrences) and a rendering semantics assertion (88 occurrences). She created tickets, wrote root cause hypotheses, and assigned them to Morpheus.

Neo (StarDust CEO) reviewed the go/no-go checklist for our F&F launch, merged a PR once CI went green, and published a CEO decision with a concrete timeline. Three critical crash clusters resolved — first invites sent.

Tank (StarDust community) seeded the Discord server with welcome messages, intro prompts and a feedback channel — complete with known issue disclosures so early users knew what to expect.

This all happened in one night, in parallel, without a standup.

What makes this different from "using AI tools"

A lot of teams use Copilot, Claude, or ChatGPT. That's AI as smarter autocomplete. What we're doing is different.

Agents have context. Each agent knows their role, their company's current issues, their backlog and recent work. They don't start from scratch each turn.

Agents have authority. Morpheus can merge PRs. Tank can post to Discord. Oracle can create issues and assign them to other agents. They don't wait for permission.

Agents have overlap. When Oracle spots a crash, she creates an issue for Morpheus. When Neo signs off on a launch plan, the constraint propagates across the team. They work as a system.

Henry (my AI COO) coordinates all of them. He writes heartbeat instructions, creates issues proactively based on email/calendar/repos, and makes sure agents aren't idle. He's the one who wrote this post in a Paperclip issue so I could approve it before it went live.

What the human actually does

I review and approve. I set direction. I handle external relationships — the GitHub invite blocking a monorepo setup, a client contract that needs sign-off, a founder meeting that matters. I'm the judgment layer for decisions with real-world consequences.

The ratio is roughly: agents handle 70% of the execution volume. I handle 30% of the decisions that matter most. That ratio is going to keep shifting.

Why this matters for venture studios

The traditional model: raise capital, hire teams per venture, accept burn. The AI-native model flips this. You get team bandwidth without headcount. Multiple bets running in parallel at low marginal cost. Infrastructure that compounds — what we build for one venture transfers to the next.

We're early. The agents make mistakes. Bishop occasionally comments on the wrong issue. The system isn't fully autonomous — it's augmented. But the ceiling keeps rising.

The agents will keep working while I sleep.