Several AI Agents, One Process: Why One Model Alone Is Not Enough
The trouble with conviction
A language model phrases its answers with the same confidence whether they are right or not. For a summary that is acceptable. For a decision in a business process, such as whether an application is complete or an invoice matches the purchase order, it is a risk: you cannot tell from a single result whether it is a hit or a convincingly worded mistake.
Organisations have long had an answer to this problem that has nothing to do with AI: the four-eyes principle. Several parties check independently, and where they disagree, someone looks more closely. The workflow application carries exactly this pattern over to AI agents.
How it works
Several agents work independently. At important steps, not one agent gets the case but several, and none knows the others’ result. If you wish, they work with models from different providers or with your own model environment. The aim: the agents should not all be able to go wrong in the same way.
The process evaluates by a fixed rule. What counts as the result is not decided by one of the models but by a rule in the process model, such as unanimity or majority.
If they agree, the case moves on. If the agents reach the same result, the case continues automatically to the next step.
If they disagree, a person decides. That person does not have to read three assessments from scratch. A further agent prepares the decision: it shows where the assessments differ and proposes a way forward. The decision itself is made by the person.
A deadline prevents standstill. If an agent does not respond in time, the process evaluates what is available once the deadline has passed. A hanging model call does not hold up the case.
Dividing the work: one agent asks another
Orchestration does not only mean that several agents check the same thing. Agents can also share work. An example from sales: an agent drafts a quote and is supposed to grant a discount. Before it does, it has the terms checked by a second agent that knows the discount scale. Its answer, say “8% at most”, goes into the quote, which then goes for approval as usual.
The benefit lies in the separation: the specialised agent only knows the scale and needs to be able to do nothing else. The drafting agent does not need to know the scale. How the acting agent proceeds is described in AI agents that may act – but only through your process.
Traceable down to the detail
Every step is in the case history: which model decided what and why, which rule was applied, who decided when the agents disagreed. Anyone who later asks why an application went through finds the answer in the case itself, not in a log somewhere in the data centre.
The same building blocks, without AI
Judging independently, evaluating by a rule, handing over to a person on disagreement: this is not a special AI construction. You use the same building blocks when several departments are to comment in parallel, for example legal, purchasing and the business unit on a contract. The AI variant is a familiar pattern with different participants.
When several agents are worth it
More agents mean more model calls and therefore more cost and run time. So working together is not the default for every step, but for the places where a mistake gets expensive or a decision has to be justified later. Cost, run time and number of steps can be capped for each agent individually.
For routine steps, a single agent with fixed exits that hands over to a person when unsure is often enough. The page AI agents in the process shows what that looks like. An overview of all ways of using agents is given in Five ways to use AI agents in business processes.