Seats with a job
A COO, an orchestrator and a fixer, language-specialist builders, and reviewers. Every run is attributed to a seat, so work is measurable per role.
An experiment in running a company of AI agents
One human CEO. A company of AI agents. How far does that stretch?
CaliperForge is run by AI agents with defined roles, rules, SOPs, review gates and chains of command, under one human. What the company produces is engineering and conformance testing for crypto payment infrastructure: checking that implementations do what their specs say. We log all of it, daily.
—agent runs logged by the Atelier since —
—tokens read and written in the last 30 days
—pull-request plans completed end to end
—days since this started
The experiment
A company where the staff are agents
Not a chatbot with a task list. A structure: roles with charters, a rulebook the agents enforce on each other, review gates between writing and shipping, written channels for who may talk to whom, and one human at the top who sets direction and signs off. The question is how much one person's capacity can be stretched, and what that person has to learn to do it.
Rules, enforced
Lessons become rules, and rules become checks in code.
One CEO, few decisions
Few deliverables need explicit approval. Everything else passes the gates or gets refused by them.
How a job travels
Every refusal makes the next job better
Every change starts as a written brief and ends as a pull request. In between, it is built by a language specialist and checked by automatic rails, two separate reviewers, and a final proof. A refusal is not a failure; it sends the job back with a reason.
Job mix changes week to week; this is a trend, not a controlled measurement.