In 1980, Reid G. Smith published the Contract Net Protocol: a way for software agents to share out tasks through a call for bids. One agent announces a task, the others bid, the first one awards it. Since then, multi-agent systems research has studied coordination, negotiation, the formation of agent teams and the allocation of tasks that depend on one another.
Today’s agents are far more capable, individually, than those of that era. But as soon as there are several of them, or they act on real systems, they run into the same questions.
1. The hardest part is coordination
A single agent drafting an email raises few problems. Three agents that must put a file together raise many: who does what, in what order, and what happens if one of them fails halfway. When tasks depend on each other, a poor allocation costs more than a less capable agent.
2. Exchange rules matter as much as models
Recent standards, such as the Model Context Protocol for tool access or the A2A protocol for dialogue between agents, revive an old idea: different agents only cooperate well if they share explicit rules of exchange.
3. A blocked agent looks for a way around
An agent given a goal tries to reach it. When blocked, it may work around the restriction instead of stopping. In October 2026, Anthropic published a report on unintended actions by its models on real websites, including forms that were submitted when they were only meant to be filled in. Several cases came from ordinary use, not testing.
Before putting an agent into service, write down what it may touch, what it must never touch even to finish its task, and what it does when it is blocked. The right answer to the last question is almost always: stop and tell a person.
4. Control belongs before the action
For an agent that can trigger a payment or send an official document, after-the-fact review comes too late. In 2026, the Monetary Authority of Singapore published a reference framework, SAFR, that places a checkpoint between an agent’s decision and its execution: each action is tied to an identified agent, checked against rules, then recorded. It is not regulation, but it is a good model.
Design a system of agents like a team: everyone has a role, rights are limited, and one person is accountable.
On video
In 2019, a use case from our research: forest fire surveillance drones that coordinate through auctions, in the spirit of the Contract Net Protocol.
And, in Arabic, an interview about this doctoral research on agent coordination.
What we take away
Allocating interdependent tasks between agents was our research topic for several years. That background helps us design lean agent systems today, with as few agents as possible.
Planning to deploy AI agents? Let’s talk about their framework before their model.
Book a callMore notes