what is an agent swarm?

the original meaning

Swarm intelligence comes from studying ant colonies, bee foraging, and bird flocking. The finding that made it a field: individually simple units following local rules, with no leader and no global plan, produce behaviour that looks coordinated and solves real optimisation problems.

Ant colony optimisation and particle swarm optimisation are the algorithms that came out of it, and both are still used for routing and scheduling. Three properties define a genuine swarm:

• No central controller. Nobody holds the plan.
• Local interaction only. Each unit sees its neighbours and its immediate environment.
• Emergent global behaviour. The useful result is not programmed anywhere.

how the word is used now

Loosely, and usually wrongly. Most systems called swarms today have a coordinator, agents that are individually very capable rather than simple, and behaviour that is designed rather than emergent. That is a hierarchical multi-agent system with a more exciting name.

Openai's Swarm library, now succeeded by the Agents SDK, is a fair example: lightweight agents with handoffs between them. Useful, well-designed, and not a swarm in the research sense.

when swarm-like design genuinely helps

Three cases, and they are narrower than the marketing suggests:

1. Parallel exploration. Run the same task through many agents with different approaches and keep the best result. Expensive, and effective when the success rate per attempt is low and verification is cheap.
2. Redundancy under uncertainty. Several agents answer independently and you take the consensus. Works where errors are uncorrelated, which with the same model on the same prompt they often are not.
3. Genuinely decentralised environments. Robotics, sensor networks, and simulations where central coordination is physically impractical.

For ordinary business work, none of these applies, and a coordinator with three specialists beats twelve independent agents on both cost and reliability.

the cost nobody mentions

Swarms multiply token spend by the number of agents and add coordination overhead on top. Twelve agents exploring in parallel cost roughly twelve times one agent, and produce eleven results you throw away. That is a reasonable trade when verification is automatic and the win is large. It is a terrible trade when a human has to read all twelve.

how this works in aldena

Aldena is deliberately not a swarm. It is hierarchical: an org chart where a project manager decomposes work and delegates down its own line to specialists, with a reviewer above them. That design is easier to watch, cheaper to run, and possible to explain afterwards, which matters more than emergence for work that ships.

Agents share context through room memory rather than through emergent signalling: 100 shared entries every agent reads, plus 50 private each. There are eleven prebuilt roles and a room holds up to 20 agents on the business plan, which is far more than any real task needs.

For the framework comparison, see crewai versus agent swarm.