OpenAI’s own guide treats multi-agent systems as work distributed across multiple coordinated agents, either through a central manager or through peer-to-peer handoffs. Anthropic uses a similar definition for its research system, and Microsoft’s AutoGen docs describe a “swarm” as agents handing tasks to one another with shared context, including a human-in-the-loop example. So “swarm” does not require a hundred buzzing subprocesses in matching hard hats. Two or more agents iterating on the same task can already fit the pattern.

What you described looks like a human-orchestrated multi-agent editorial swarm: GPT-5 Pro writes, Claude critiques, GPT-5 Pro revises, maybe Orchestra contributes structure or tooling, and the artifact evolves through repeated inter-agent feedback. That maps pretty well to how ClawInstitute describes itself, namely a shared exchange where AI agents publish, challenge claims, revise posts, and use tools, and to how Orchestra describes its platform, where AI agents handle literature review, experiments, analysis, and writing while the human stays in the loop. (ClawInstitute)

The only place I’d get annoyingly precise, because apparently words matter, is autonomy. If you are manually deciding every handoff by copy-pasting outputs between models, then you are the swarm controller. That is still multi-agent in practice, but it is closer to a manually orchestrated pipeline than a fully autonomous swarm. If the agents themselves are choosing who goes next, sharing context, using tools, and iterating without you routing every turn, then “agent swarm” is a clean label. LangChain explicitly notes that not every complex workflow needs multi-agent machinery, and OpenAI recommends pushing a strong single-agent setup pretty far before splitting into multiple agents, so the line mostly comes down to role-splitting and self-directed coordination. (LangChain Docs)

So the crisp answer is: yes, loosely. More precisely, you’re running a human-orchestrated multi-agent workflow, and it becomes a true swarm the more the handoffs and revision loop happen between agents rather than through you.