About

Why Sumn exists.

Agents got good at real, recurring work before businesses could trust them to run it unattended. Sumn is the control plane that closes that gap. You set the rules once. Sumn enforces them on every run and reworks the result until it passes review or hands it back to you.

Every team that uses agents for real work lands in the same place. The work gets done, but a person babysits it the whole way, re-prompting the agent and moving its output from one window to the next. The agent is capable and the job repeats, and a human is still the glue and the safety net. What agents can do for the business is capped at the hours that person can sit there.

What we believe

A smarter model won't fix this. Structure will. Most guardrails for agents are instructions in a prompt, and the model is asked to stay in bounds. Sumn puts the bounds in the runtime instead, so staying inside them isn't left to the model. It still makes every judgment call, inside limits it can't talk its way past.

Spend is the clearest case. Sumn checks the cap before each model call. A usage dashboard reports what was spent after the fact, so an agent stuck in a loop can run up thousands overnight before anyone looks. Because the check comes first, the run stops at the number you set.

No model is best at everything, and that holds inside every lab. Model skill is jagged: the one that's brilliant at a hard judgment call is mediocre at a routine one. So on a playbook, the model for each stage is just a setting. Put a cheap open-weight model on your own keys where the work is routine, and a frontier model where the judgment earns its cost. Change either the day a better one ships or a lab goes down.

This only works if the platform stays neutral. Sumn doesn't sell models. It runs the ones you choose, open-weight seats on your own keys, and takes no side in which lab wins. A neutral control plane can't come from a token vendor.

How we work

We're our own first customer. The code, research, and content behind Sumn move through the same kind of playbooks we sell: gated agent pipelines, several models to a job, and adversarial review at the gates. This page came through one of them. Running it that way every day is how we find the rough edges before a customer does.

Where it's heading

Playbooks come first. A playbook is a recurring job written down the way you'd brief a capable person: its stages, what each one may spend, who or what signs off before the work moves on. The company that runs it owns it and versions it, the way it would any process it depends on. A canned agent from a lab is the other thing entirely: you rent it, and it changes on the vendor's schedule. Playbooks are what you can run on Sumn today, and more will sit on top as the platform grows.

Each playbook is one component of a company that runs itself. String enough together and a business runs its recurring work without a person steering every run. Keep going and you reach the one-person, billion-dollar company: a founder whose operations are playbooks instead of a payroll. The infrastructure that company runs on is what we're building.

We're onboarding a small number of early teams now. Tell us the process you'd run first, or write to hello@sumn.run.