You can't fire an agent

Half the founders I talk to are building multi-agent systems. Some spin up agents for breakfast. A few, Jason Lemkin among them, write genuinely good long-form about what they are learning, and I read all of it.

We went the other way. We run three agents.

One agent inside the product

Simba, our AI agent, is embedded everywhere in Etlworks. It is the vibe coding agent, the monitoring and ops agent, the Q&A agent, the marketing agent, the data analytics agent, the code review agent. Pick your favorite specialization. It is the same agent holding a different tool.

That is the whole trick. An agent is not defined by its personality or its system prompt. It is defined by what it can reach. Give one agent real tools and real permissions and it covers the ground that a dozen wrappers were invented to cover.

Two agents outside it

Claude Code and Codex. That is the rest of the list.

They write our code. They refine our documentation. They update this website. They work on copy. They handle a growing share of the everyday business that used to eat entire afternoons.

The uncomfortable part, stated plainly: those two are better than a thousand specialized agent wrappers, and it is not close. The wrapper industry exists because general agents were weak eighteen months ago. They are not weak now.

They also build things that are not agents

This is the part people skip. Our agents spend much of their time writing deterministic tools.

Cloud cost monitoring. Extracting marketing signals from a dozen sources. Keeping infrastructure online. Dozens more. These are programs, with repos and commits and tests, deployed wherever we need them: our cloud, local laptops, customer environments.

They are not prompts. A prompt running in production is a coin flip you have automated. A program does the same thing every time, and when it is wrong you can read it and find out why.

Ownership is the actual argument

I believe in this technology. I do not believe in an over-engineered mesh of agents with no clear owner.

Every piece of work here belongs to a person. When something goes wrong I ask that person what happened. I get an answer. It gets fixed.

I do not want to open an investigation into which of my forty agents made the call. I do not care which one did. I cannot fire an agent. An agent does not care that it was wrong, will not lose sleep, and will not do better next quarter because of the conversation. Accountability is not a feature you bolt onto a swarm afterwards.

Humans own the work. Agents do the work. That distinction carries more weight than any orchestration diagram.

The cost question, since everyone asks

We are roughly 10x more productive than we were in 2024.

We do not count agentic tokens and we do not cap them. Nobody here has been told to use less. And we spend less than we did in 2024.

If your agent architecture needs a diagram, a budget committee, and an on-call rotation for the agents themselves, it is worth asking what all of it is buying you.