I built a tiny AI company that runs on my laptop (multi-a...

Lately I’ve been thinking about what “work” might look like in a world of AI agents.

I was talking with my wife about it — maybe in the future, instead of replacing people one by one, it becomes one person managing a small team of AI workers.

But tools like Claude Code or OpenClaw felt way too technical for her.

So I started wondering:

what if this actually felt like working with a team?

Not writing code.

Not wiring workflows.

But hiring, managing, and collaborating.

So I built Holons.

What it looks like

Holons is a desktop-first system (Tauri + Flask + React) where:

agents have roles and identities
a “lead” agent assigns tasks and builds workflows
multiple agents can collaborate in group chats
everything runs locally (or self-hosted)

You can say something like:

“Create a pitch for a B2B AI accountant”

And it will:

propose a workflow
assign tasks to different agents
estimate cost
and let you run the whole thing

What surprised me

The hardest part wasn’t the agents.

It was visibility.

Once you have multiple agents running, you need to understand:

what they’re doing
how much they cost
how they behave over time

So one design decision I made early:

Every LLM call writes to a unified ledger

Each call tracks:

model + provider
tokens (prompt / completion)
cost
latency
agent + user

That single table powers:

cost dashboards
usage quotas
reports
audit logs

This turned out to be way more important than I expected.

Some technical notes

Multi-provider support: Bedrock, OpenAI, Anthropic, Gemini, MiniMax
Per-agent model binding (different agents can use different models)
pgvector-based RAG + external knowledge integrations
MCP-style tool integration
IM channels (Telegram / Slack / LINE)

Still early

This is very much an experiment.

I’m not sure yet if this is the “right” abstraction for multi-agent systems,
but it already feels different from traditional workflows.

More like managing a team,

less like calling an API.

Links

Curious how others are thinking about multi-agent systems.

Are people actually using them in real workflows,
or are we still figuring out the right model?

推荐订阅源