This page covers five open-source frameworks at the round-up tier: a paragraph of framing, a paragraph of when-it-is-right, a paragraph of when-it-is-not, and a short operator note. Each could justify a full review; for v1 we are publishing the round-up because the use cases overlap meaningfully and the cluster is more useful as a comparison than as five separate pages.
For the four frameworks we cover at full-review depth, see LangChain, LangGraph, CrewAI, and AutoGen.
AutoGPT
The OG. The 2023 viral moment that proved autonomous agents could go viral. In 2026, mostly retrospective interest.
When it is right: Pedagogical use. Reading the AutoGPT codebase is a useful way to understand the autonomous-agent loop in its earliest production-adjacent form.
When it is not: Production. The architecture has been superseded by every commercial autonomous tool and by OpenClaw. Use the lesson, not the code.
Operator note: We ran AutoGPT in 2024. The cost-per-task variance was extreme. The Cost Cliff failure mode was named retrospectively against it.
MetaGPT
Role-based multi-agent that predates CrewAI's mainstream moment. Originally framed as a software-company simulation.
When it is right: Research-shaped multi-agent prototypes. The role-based abstraction is clean for academic-style problems.
When it is not: Production at scale. Like CrewAI, the role-coordination overhead grows with agent count.
Operator note: MetaGPT remains active; the project shape is closer to research than to production-ready framework. Treat it as such.
Pydantic AI
Type-driven agent framework from the Pydantic team. Strong typing, structured-output discipline, deliberate scope.
When it is right: Production agents where structured-output validation is the dominant concern. The Pydantic-typed I/O is a real advantage when you need it.
When it is not: Multi-agent orchestration. Pydantic AI is single-agent-shaped; for orchestration use LangGraph.
Operator note: We use Pydantic in production and we like Pydantic AI for the structured-output use case. For pure orchestration, LangGraph is the better fit.
DSPy
Compile-time prompt programming. Define a signature, optimise the prompt automatically against examples.
When it is right: Tasks where prompt engineering is the bottleneck and you have a labelled dataset to compile against.
When it is not: Tasks where the prompt is already short and the dataset is unavailable. DSPy can be over-engineered for simple tasks; use the cheaper pattern first.
Operator note: DSPy is genuinely useful when the conditions match. The conditions match less often than the project's enthusiasm suggests.
smolagents
The minimal one. Hugging Face's small agent framework with a deliberate focus on clarity over features.
When it is right: Educational deployments and simple agents where you want to read every line of the framework.
When it is not: Complex orchestration or production scale. The minimal scope is the feature; do not ask the framework to do more than it is designed to do.
Operator note: We like smolagents for prototyping and for teaching. We do not run it in production.
Oliver runs Digital Signet, a research and product studio that operates ~500 production sites with AI agents as the engineering layer. The Digital Signet portfolio is built using a continuous AI-agent build pipeline, one of the largest agent-operated publishing operations on the open web. The handbook draws directly from those deployments: real cost data, real failure modes, real recovery patterns.