Prompt Engineering in July 2025: Async Agents, SmolLM3, and the Next Wave of LLM Innovation

The world of prompt engineering and large language models (LLMs) continues to evolve at a breakneck pace. This week, several major developments from OpenAI, Hugging Face, and Google AI are shaping the future of how we interact with and deploy advanced AI systems.

Hugging Face: Async Agents and SmolLM3

#

Hugging Face has been on a roll this July, publishing a series of articles that push the boundaries of what LLMs can do. The standout is their work on asynchronous robot inference—decoupling action prediction and execution for more efficient AI agents (read more). This approach is crucial for real-world applications where latency and parallelism matter, such as robotics and autonomous systems.

Another highlight is SmolLM3, a small, multilingual, long-context reasoner (read more). SmolLM3 demonstrates that you don’t need massive models to achieve impressive reasoning and multilingual capabilities. For prompt engineers, this means more accessible, efficient models for specialized tasks and languages.

Hugging Face also released guides on building MCP (Model Context Protocol) servers and deploying full-stack desktop agents, signaling a shift toward more modular, composable AI systems.

OpenAI: New Collaborations and Model Safety

#

OpenAI’s latest news includes a high-profile collaboration between Sam Altman and Jony Ive (read more), hinting at new directions in AI product design and user experience. While details are still emerging, the partnership underscores the growing importance of human-centered design in prompt engineering.

On the technical front, OpenAI continues to focus on model safety, responsible disclosure, and the rollout of new models like o3 and o4-mini. Their recent releases emphasize the need for prompt engineers to stay vigilant about security, privacy, and ethical deployment as LLMs become more deeply embedded in business and society.

Google AI: Graph Foundation Models and Health AI

#

Google AI’s July updates feature the introduction of graph foundation models for relational data (read more), which could revolutionize how LLMs understand and reason about complex relationships. This is especially relevant for prompt engineers working on knowledge graphs, recommendation systems, and enterprise data integration.

Additionally, Google’s MedGemma open models for health AI (read more) highlight the trend toward domain-specific LLMs. For those crafting prompts in regulated or specialized industries, these advances open new possibilities for safe, effective AI deployment.

Key Takeaways for Prompt Engineers

#

• Async and Modular AI: Asynchronous inference and modular architectures are making AI agents faster and more adaptable.
• Smaller, Smarter Models: SmolLM3 and similar projects show that efficiency and multilingual support are within reach for smaller teams and organizations.
• Human-Centered Design: Collaborations like OpenAI’s with Jony Ive point to a future where prompt engineering is as much about UX as it is about tokens and context windows.
• Safety and Ethics: With new models come new responsibilities—security, privacy, and ethical considerations must be top of mind.
• Domain-Specific LLMs: From health to enterprise data, specialized models are unlocking new prompt engineering challenges and opportunities.

As always, the best way to keep up is to stay curious, experiment with the latest tools, and share your findings with the community. The prompt engineering revolution is just getting started.