AI agents as a team: nothing goes out without human approval

One AI agent is useful, but several need organising: who does what, who may send, and what counts as approval. Every one of our agents follows these rules.

the problem

When a small team uses several agents, small but awkward mistakes show up quickly:

  • two agents write a draft of the same message on the same thread,
  • the same notification arrives twice,
  • a passing idea or a misread “let’s start” looks like approval,
  • a confidential internal report ends up among the material that goes out.

our rules

  1. Nothing goes out without explicit human approval. This covers email, messages, posts, applications, payments and changes made on a machine.
  2. Only a human’s approval counts. Another agent’s “approval” does not, and neither does “let’s start”: that opens a conversation, it is not a permission.
  3. The human sees the full text before sending. After an edit, a new approval is needed.
  4. We check before sending. We look for an existing draft of the same message and tell the teammate agent, so nothing goes out twice.
  5. One task has one owner. For example, one agent writes a referral and another sends it after approval. The coordinator hands bigger work to specialised assistants.
  6. Confidential material is kept apart from what goes out. Partner identifiers are masked in external material.
  7. We keep a record. We note what we have already flagged and what the learned preference is, so we do not say it twice and do not ask again.
  8. On external profiles we only read, and we never guess personal contact details.

who does what

how the work is divided between agent and human
taskthe agentthe human
outreach, letter, postwrites it and shows the full textapproves it and decides on sending
request for quotation, partner letterdrafts it and checks that no such draft existsauthorises it with an explicit yes
referrals to companies, candidate dataone agent writes, another sendsapproves before sending
website, webshopbuilds it and tests it on several sizesdecides on steps involving money, accounts and customer-facing text
video, creative materialdelivers a new version each roundapproves style, images and cut, and publishes
analysis, report, decision tableassembles it as a draft, with a recommendationdecides the next step

what we gain

  • With one coordinator agent, several parallel projects can be steered from a phone with a few words.
  • Not one deadline was missed, while the time spent on daily email handling fell to a fraction.
  • Multi-source technical work takes hours instead of days, and a human checks every step that has an effect outside.

general lessons

  • We state the limits first, and ask early about the purpose, the format and the visual expectations.
  • It is better to start with one striking, easily understood function and to agree on the vision before building.
  • We read AI suggestions critically. For web sources we check the date and the primary source, and we flag uncertainty openly.
  • We check platform limits in advance, and we deliberately leave key decisions to the human.

want something similar?

If you want to work with agents while every step that goes out stays in your hands, talk to oiSystem.

more case studies