Swarm, a new agent framework by OpenAI

Original Article ↗ Hacker News Discussion ↗

Scope, Licensing, and Intended Use

  • Swarm is MIT-licensed and positioned as an experimental “sample framework” for multi-agent systems, explicitly not production-ready and not officially supported.
  • The repo states PRs and issues will not be reviewed, which some see as logically separate from “not for production” and effectively discouraging collaboration.
  • A linked cookbook example exists, but people note basic spelling/grammar errors in official OpenAI content and question whether OpenAI actually uses its own models for documentation.

Design, Code Quality, and Comparison to Other Frameworks

  • Some find the code “poorly written” (no async, heavy deepcopy, print debugging) and see it as a simple reference rather than a serious framework.
  • Others argue that’s acceptable given it is explicitly a sample/experimental library.
  • Multiple alternatives are suggested: LangChain, LangGraph, Autogen, txtai, Langroid, Microsoft Semantic Kernel, crewAI, griptape, and others.
  • One view is that what LangChain-style frameworks do is simple enough that many teams just roll their own instead of adopting yet another abstraction.

Multi-Agent Orchestration & Technical Challenges

  • Several commenters claim Swarm offers nothing fundamentally new versus many existing agent frameworks.
  • A recurring theme: the “hard part” is not routing/triage of prompts but:
    • Handling long-running, large-compute inference with robust message-passing.
    • Dealing with high-bandwidth, multimodal data between many agents.
    • Designing and optimizing agent graphs and workflows rather than a single prompt.
  • There’s debate over infrastructure choices (Temporal, Kafka, etc.), with some dismissing them as reinventions of older ideas (e.g., Erlang-style systems).

Production Use, Reliability, and Hype

  • Multiple people question whether multi-agent systems are actually working at scale in production, citing slowness, cost, and unreliability.
  • Others report real use cases:
    • Internal batch agents for large-scale code generation and testing.
    • Support-fraud analysis systems where cost and latency are secondary to accuracy and analyst assistance.
    • Personal agents used daily for research and data analysis.
  • A key problem raised is “divergence”: ensembles of agents drift from goals, requiring strong constraints and ground-truth checks.
  • Some argue that rapidly improving large-context models and newer APIs may make complex agentic setups less necessary; others counter that evals on real tasks still show benefit from carefully designed workflows.

Naming, Trademark, and Ecosystem Drama

  • The name “Swarm” clashes with:
    • A 1990s multi-agent simulation toolkit.
    • A separate, heavily-promoted “Swarms” agent framework whose author has been criticized elsewhere for low-quality or non-functional repos.
  • There is an ongoing trademark complaint around “swarms”; several commenters think it is unlikely to succeed given long-standing generic use of the term.