Swarm, a new agent framework by OpenAI
OpenAI’s release of Swarm, an MIT-licensed but explicitly experimental framework for orchestrating multiple AI agents, is prompting mixed reactions from developers. Some see it as yet another lightweight, non-production-ready toolkit in an already crowded space alongside LangChain, Autogen, and others, criticizing its code quality and questioning whether complex agent architectures are necessary given rapid model improvements. Others focus on deeper issues: whether multi-agent systems can be reliable, cost-effective, and genuinely useful at scale, while side debates touch on vendor lock-in, naming conflicts with existing “Swarm/Swarms” projects, and the broader hype vs. reality of AI agents in real-world products.
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.