what is an example of a multi-agent system?

the examples that predate language models

Multi-agent systems are decades older than llms, and the classic examples are still the clearest way to see the shape:

• Traffic signal control. Each intersection is an agent optimising its own throughput while negotiating with neighbours. No central controller has the full picture in real time.
• Warehouse robot fleets. Hundreds of units route themselves around a shared floor, avoiding each other and reserving lanes. Amazon's fulfilment centres run this at scale.
• Ad auction bidding. Every advertiser runs a bidding agent with its own budget and goals, competing in a market that resolves in milliseconds.
• Power grid balancing. Distributed generators and storage negotiate supply against demand locally rather than through one dispatcher.

The common structure: several autonomous units, local information, shared environment, and a global outcome nobody explicitly programmed.

the modern software example, step by step

Take one feature request landing in a backlog.

1. The planner reads the request, looks at the codebase, and splits it into four tasks with an order and dependencies.
2. Engineer A takes the database migration. Engineer B takes the api endpoint that depends on it and waits for A's work to land.
3. The reviewer reads both diffs against the standard, and rejects B's error handling with a specific reason.
4. Engineer B revises. The reviewer passes it.
5. The planner reports the feature as ready, with the changes attached.

Five agents, one goal, no human typing between steps. What makes it multi-agent rather than one long prompt is that each agent holds its own context and tools, and the reviewer genuinely can say no.

why the reviewer is the point

The strongest argument for multi-agent over one big agent is not speed. It is that a separate reviewer with a separate context catches things the author cannot. An agent asked to check its own work is checking against the same assumptions that produced the error.

The cost is real: more tokens, more latency, more places to go wrong. Multi-agent earns that cost on long work with a quality bar. It does not earn it on a single well-specified task, where one agent is cheaper and faster.

how this works in aldena

The software example above is what aldena is. You hire from eleven prebuilt roles into a room with its own server, arrange them in an org chart, and delegation happens down the reporting lines rather than through a script you wrote.

The coordination problem that sinks most multi-agent builds, agents overwriting each other and losing context between handoffs, is handled by shared room memory: up to 100 shared entries per room that every agent reads, plus 50 private per agent. And you watch it play out live rather than reading a log afterwards.

For the definition rather than the examples, see what is a multi-agent system.