Context card
Agent fleet or agent swarm: what parallel activity shows
What is the difference between an agent fleet and an agent swarm, and which of the two can an outside observer establish?
The short answer
An agent fleet is a number of AI agents that run similar tasks in parallel without exchanging information. An agent swarm is a group of agents that exchange messages and adjust to each other, so the group behaves in a coordinated way. Observations made from outside — addresses, hosting, request patterns, timing — normally support only the weaker description, the fleet.
For: Security and documentation staff assessing agent activity reports
Key points
- "Fleet" claims quantity and similarity of tasks.
- "Swarm" claims an information path between the agents and mutual adjustment.
- Shared hosting shows where traffic came from, not who commissioned it.
- Identical request patterns can come from one widely copied template.
- The stronger claim needs records from inside the runs.
The context
Two descriptions, two different claims
Calling a group of agents a fleet asserts little: many instances, similar tasks, parallel operation. Calling it a swarm asserts more. A swarm needs communication between the agents — messages, a shared task store, a dispatcher — and it needs the members to change their behavior because of what the others do. The second claim is therefore harder to support, and conflating the two makes an observation sound like a finding about intent.
What an outside observation carries
From outside, an observer sees requests arriving at a target, entries in passive scan or certificate data, address ranges in a provider's network, user agent strings and timing. Those traces are real but thin. Hosting is rented and resold, so an address range identifies an infrastructure, not an operator or a sponsor. Similar prompts and request sequences circulate as templates, so many unrelated users can produce near-identical traffic. Volume, finally, says nothing about purpose: mapping queries, price checks and reconnaissance look alike at the transport level.
What would make the stronger claim checkable
Coordination becomes visible in the records of the runs themselves: the task that was issued, the identity each agent used, the permissions in force, the sequence of tool calls and any message traffic between the agents. An operator of its own fleet can produce this; an outside observer usually cannot, and a security report should say which of the two positions it is written from.
For technical documentation the distinction has a practical use. If a product runs many agents, the documentation should state whether they are independent instances or whether a coordinating component exists, where it is, who may configure it and what happens when it fails. Readers who have to size rate limits, support contracts or incident procedures need that answer, and the words "fleet" and "swarm" are not interchangeable in it.
Questions readers ask next
- Is a dispatcher that distributes tasks already a swarm?
- It is central coordination, not mutual adjustment between peers. Many systems sit between the two: a planner splits work, the workers never talk to each other. Describing the actual topology is more useful than the label.
- Why does the distinction matter for a risk assessment?
- A coordinated group can pursue a goal across targets and adapt to defenses. Independent parallel runs mainly produce load and repeat the same mistake, which calls for different countermeasures.
- Can traffic be assigned to a particular model provider?
- Rarely with certainty from outside. Request patterns and user agents can be set freely, and traffic may pass through intermediaries, so such an assignment needs corroboration from the provider or the operator.
Sources
- Agentic AI – Threats and Mitigations — OWASP GenAI Security Project, October 13, 2026
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 — NIST, January 26, 2023
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.