Making several AI agents work together, without chaos
One role per agent, written hand-overs, shared rules, a single source of truth: how to organise several AI agents like a team.
A first agent is a project. Three agents are an organisation. And like any organisation, they need roles, rules and hand-overs, otherwise each one works in its own way.
What happens without organisation
- Contradictory answers: the customer service agent announces a delivery time the logistics agent does not know about.
- Duplicates: two agents write to the same customer on the same day.
- Orphan cases: a request goes from one agent to another, and nobody finishes it.
- Rules that diverge: each agent has its own copy of the prices, and they are no longer updated at the same time.
An organisation in levels
The simplest approach is to model the agents on your organisation:
| Level | Role | Example |
|---|---|---|
| Shared memory and rules | what applies to the whole company | priorities, vocabulary, trade-off rules |
| Business agents | a whole business domain, with its knowledge | purchasing, customer relations |
| Process agents | a process, end to end | order processing, qualifying incoming requests |
| Tasks | a specific action | sorting incoming emails |
Each agent knows where its rules come from, and to whom it reports.
The five rules that prevent chaos
1. A clear role for each agent
What it does, what it does not do, to whom it hands over. Two agents must not have the same role on the same case.
2. Written hand-overs
When a case passes from one agent to another (from qualifying a request to handling it, from the order to the invoice), the hand-over is described: which information, which format, who checks.
3. Shared rules, in a single place
Prices, lead times, vocabulary, priorities: a single source, which all agents refer to. Two copies always end up diverging. See Company memory: five layers to make it last.
4. Decisions reserved for people
The more agents there are, the clearer you must be about what remains human: paying, signing, setting a price, publishing, settling a strategic decision. See Governance.
5. Records that follow the case
Being able to reconstruct a case’s journey, from one agent to another: that is what makes it possible to understand an error and correct it.
In a mid-sized company: consistency before numbers
In a mid-sized company, agents often arrive department by department, with different providers. The risk is not the number of agents: it is the inconsistency between them. Writing shared rules and governance that apply to all agents costs less than reconciling agents that contradict each other. See From AI pilot to production in a mid-sized company.
Where to start
- One process, then others if they stay independent: the Process Agent.
- The whole business domain, as soon as several processes share its rules: a Domain Audit, required, then Business Agents, with its knowledge written down for the agents.
- Shared rules and memory, when several business domains are involved: the Company Brain.
You pay for a result, not for a number of agents: how many are needed is our business.