The agents that run this company.

What this does

A set of agents, each with one defined job, running on schedule and on trigger. A morning brief that reads the sources so the day starts informed. Inbox triage that separates what needs a person from what does not. Draft production. Uptime and error sweeps across everything Modulus runs.

Each one produces a specific, checkable output and writes a line to a log. Nothing runs invisibly, and nothing that leaves the building does so without a person approving it.

See it work

1. A trigger fires A schedule, a webhook, a new file, an inbound email.
2. One agent picks it up One job, one agent. No general-purpose assistant improvising across six responsibilities.
3. It calls the tools it needs Search, the file store, the calendar, the database — scoped to that job and nothing wider.
4. A human check where it matters Anything that reaches a client, a customer, or the public waits for approval.
5. The run is logged Time, job, outcome. When something goes wrong you can see exactly where.

Why I built this

One person cannot be a research department, a content department, and an operations department at the same time. Something gives, and it is usually the work that has no deadline attached — which is, inconveniently, most of the work that compounds.

But I was not willing to sell agents to a small business on a theory. Small businesses do not have a spare quarter to donate to my learning curve. So the fleet exists to take the losses first: to find out which jobs an agent can actually hold, which ones drift after two weeks, and which ones were never a good idea in the first place.

What I learned is unglamorous and it shapes everything I install. An agent with one narrow job and a measurable output is dependable. An agent asked to ‘handle marketing’ is a liability with a friendly tone.

How it is built

• n8n for orchestration The workflow is a diagram you can read, not a script only I understand. If Modulus vanished tomorrow you would still be able to open it.
• Claude by Anthropic for judgment Used where judgment is genuinely needed, and nowhere else. Deterministic steps stay deterministic.
• Its own locked-down machine The fleet does not share a box with anything else, and it does not run on a laptop that goes to a coffee shop.
• Secrets never live in the workflow Credentials sit in a manager and are injected at run time, so exporting a workflow never exports a key.
• Every run logged Because ‘it seems to be working’ is not a status anyone should accept.

Where this breaks

This is the section most people selling agents leave out.

• A vague job produces a drifting agent ‘Keep an eye on the competition’ will look fine for a fortnight and then quietly go strange. The fix is not a better model; it is a narrower job with a defined output.
• If you cannot measure the output, you cannot own it Any job with no checkable result does not go to an agent. It stays with a person until someone can say what ‘done correctly’ looks like.
• Upstream outages stop the fleet When a provider goes down, the affected jobs stop. They queue and resume rather than half-finishing, which is the correct behaviour and still an outage.
• It does not remove the human It removes the waiting. Anything that reaches a customer still passes a person first, by design and not by accident.

Build a workflow like this

We start with one job — the specific piece of your week that is repetitive, defined, and currently waiting on you. Not a platform. Not a transformation. One job, running, measured, and owned. If the first one does not pay for itself, there is no second one.

Book my free 20-minute call