DSPy
A Stanford-built framework that treats LLM prompts as compilable programs — define what you want declaratively, DSPy optimizes the prompts and few-shot examples automatically.
DSPy flips the prompt-engineering workflow. Instead of hand-writing prompts and tweaking by feel, you define a "signature" (input/output schema) and DSPy compiles the prompt by sampling examples, evaluating against metrics, and selecting the best configuration. The result: production-quality prompts without the manual iteration.
For sophisticated production teams, DSPy is increasingly the right approach for any LLM-driven step with measurable success criteria. The learning curve is steeper than direct prompt engineering, but the optimized prompts often outperform hand-tuned alternatives.
In 2026, DSPy is most popular in academic and AI-research-heavy startups. For typical production use, the simpler vendor-SDK + manual prompting approach is still more common. Expect DSPy adoption to grow as more teams hit prompt-quality ceilings.
Frequently asked
When should I use DSPy?+
When you have an LLM step with measurable success criteria and you are hitting the ceiling of hand-tuned prompts. For early-stage prototyping or simple agents, direct prompting is simpler.
Is DSPy production-ready?+
Yes for the right use case. Used by several production AI startups and major enterprises. The learning curve is real but the optimization gains justify the investment for mature LLM teams.