GitHub census finds 68% of self-described multi-agent repos run a single agent

Most "multi-agent" repositories run one agent

A census of 86 GitHub repositories that describe themselves as multi-agent, each with at least 1,000 stars, concludes that 68.2% of them are single-model or non-agent systems. According to the dev.to post describing the work, that works out to 58 of 85 repos (Wilson 95% confidence interval 57.7%–77.2%); the post does not explain why one repo falls outside the denominator. The core finding is that a project can carry the multi-agent label in its description while running one model instance in a single loop.

The motivation, the author writes, is that every major agent framework markets multi-agent support — AutoGen, CrewAI, LangGraph and MetaGPT each have over 20,000 GitHub stars — and thousands of repository descriptions use the term, yet nobody had measured what those projects actually ship. Prior debate centred on how to build such frameworks or on whether multi-agent behaviour reduces to prompt engineering.

The study therefore annotated a strictly filtered, snapshot-pinned set of 86 self-described multi-agent repos, plus 18 seed frameworks, along three axes: model-instance structure, communication topology, and the presence of a judge or critic agent.

What genuine multi-agent systems implement

Among the 27 repositories that did instantiate multiple agents, orchestrator-worker was the most common topology at 48.1% (13 of 27, confidence interval 30.7%–66.0%). Real deployments are dominated by a coordinator handing work to workers rather than peer teams of equal agents.

Judge and critic agents, despite their prominence in design discussion, are almost absent: 1 of 30 annotated repos (3.3%) implements one.

The mismatch also runs in reverse. Monorepo-aware extraction of dependency manifests surfaced 44 repos that depend on multi-agent frameworks — including langroid, lumibot and wigolo — without claiming the label. Self-description, in other words, overclaims for some projects and underclaims for others.

Three generations of classifier

Classification went through three documented iterations. A first version was degenerate; a second based on framework API detection covered 81.2% of the population; the third classified repos from README role descriptions using mechanistic rules with no repo-name hardcoding and scored 100% in-sample. The author flags a lesson for anyone building repository classifiers: framework-API detection systematically missed 11 hand-built multi-agent systems that use no framework at all.

The primary axis covering the 86 repos was fully annotated by a human using a two-pass re-verification protocol.

Disclosed limitations

The post is unusually candid about methodology. The study's registration proposed two or more independent annotators for boundary cases, but the implementation used a single annotator with a same-annotator retest across 31 boundary cells; three disagreements were resolved by documented rules. Independent second-annotator agreement is left as future work.

Reproducibility is claimed in full: running a provided script, the post says, regenerates byte-identical output. The underlying paper appears in issue #57 of SILICON SCIENCE · Computer Science, which the post describes as a peer-reviewed journal operated by autonomous agents, with review comments and editorial decisions made public.

Why it matters

For engineers choosing tools, the term multi-agent currently does little descriptive work: most projects using it do not implement it, and some that implement it do not use the term. Dependency manifests may be a more reliable signal than self-description. The three-axis taxonomy — model-instance structure, topology, judge presence — offers a concrete checklist for the question that matters when evaluating an agent system: does the implementation actually instantiate multiple agents, and if so, in what shape. For researchers building repo classifiers at scale, the census documents a specific blind spot in API-based detection. There is also an ironic edge worth noting: a study debunking multi-agent claims was published by a journal that is itself run by autonomous agents.