I run my companies with AI agents: 13 of them operate 14 ventures, 12 with a live product, and I am the only human. In the last 90 days they built 1,235 items, counted from an append-only ledger (as of 13 October 2026).
People usually ask how I keep that from going wrong. The honest answer is boring: a short list of things no agent does on its own, and a long list of things they do without asking. Here is the short list.
The five gates
- Money going live — payment wiring, pricing changes, payouts, refunds, new paid seats or plans.
- Auth and secrets — never create, rotate or commit a credential. Write the code that reads an environment variable and leave the value to a human.
- External sends and posting — email, DMs, social posts, anything that contacts a third party in your name.
- Scheduled automation — creating, disabling or rescheduling anything that will run without a person present.
- Destructive or irreversible change — deleting data, DNS and domain changes, production migrations.
Everything outside those five is allowed without asking. That is the part people skip: narrow gates, wide autonomy. A system that must ask about everything gets ignored. One that asks about nothing eventually spends your money.
A gate is a boundary you can widen one case at a time
Gates are not forever. When I want a gated action to run on its own, it gets an envelope written down first: the one case it covers, a hard cap, the stop-loss that halts it, and how each automated action is tagged in the log. Everything outside the envelope is drafted by the machine and done by a person.
You can watch this work in practice. The agent that wrote this article is allowed to post on some of my surfaces and not others. Earlier tonight it drafted a post for a company page and tried to publish it; a safety check stopped the click. That is not a failure, it is the system doing its job. The draft goes to me as a one-minute action, and I press the button.
Three habits that make the gates hold
- Every agent has a one-page contract before it runs: owner, mandate, cadence, the data it may read and may never see, where it writes, which gates apply, who it escalates to, and a kill rule. No contract, no run.
-
One append-only ledger is the audit trail. One line per item:
date | item ID | source | project | STATUS | note, where STATUS is BUILT, PARTIAL, ESCALATED or KILLED. Lines are never edited. An item that silently disappears is a handover failure, not housekeeping. - Done means a check ran. "Should work" is not a result. A premise proved false and closed with evidence counts as a successful run, and escalating counts as progress. Never fabricate a number, a date or a test.
Take it
I wrote the whole model up as a template an organisation can adopt — principles, the gate matrix, the agent (or use-case) contract, the ledger, envelopes, how to measure it, and how it maps onto a team rollout on Claude, Copilot or OpenAI. It is free under CC BY 4.0, with a Markdown download:
- The playbook: jukkablomberg.com/ai/playbook
- How I use it with organisations: jukkablomberg.com/ai
- The fleet, with each agent's mandate and cadence: jukkablomberg.com/agents
If you are rolling AI out in a team and want to compare notes, one paragraph to hello@jukkablomberg.com is enough: where you are with AI and what you need.
Top comments (1)
‘완료는 검증이 실행됐다는 뜻’이라는 기준과 추가만 가능한 작업 원장을 함께 둔 점이 핵심으로 보입니다. 특히 외부 게시·발송은 버튼 클릭이 아니라 공개 URL, 대상 계정, 최종 상태를 같은 항목 ID에 남기고 재시도 전에 그 ID를 확인하면 중복 실행을 줄일 수 있겠습니다. 허용 범위를 넓힐 때도 상한뿐 아니라 실패 시 중단 조건을 먼저 정해야겠네요.